I would like to understand the "justification" for the bilinear transform. The basic idea as I understand it is that by integration rule of Laplace transform we have for continuous $y(t)$:

$$\mathcal{L}\{y\}(s)=\mathcal{L}\{\int_0^ty'(x)dx\}(s)=\frac{1}{s}\mathcal L\{y'\}(s)$$ $$\Leftrightarrow \frac{\mathcal L\{y\}(s)}{\mathcal L\{y'\}(s)}=\frac{1}{s}$$

Using the trapezoid rule gives the approximation ($h$ is the sampling period): $$y(t)=y(t-h)+\int_{t-h}^ty'(x)dx\approx y(t-h)+\frac{h}{2}(y'(t)+y'(t-h))$$

And taking the $z$ transform using the approximation gives $$\mathcal Z\{y\}(z)\approx \mathcal Z\{ y[n-1]+\frac{h}{2}(y'[n]+y'[n-1])\}(z)$$ $$\approx z^{-1}\mathcal{Z\{y\}(z)}+\frac{h}{2}(\mathcal Z\{y'\}(z)+z^{-1}\mathcal Z\{y'\}(z))$$ $$\Leftrightarrow\frac{\mathcal Z\{y\}(z)}{\mathcal Z\{y'\}(z)}\approx \frac{h}{2}(\frac{1+z^{-1}}{1-z^{-1}})$$

Putting the above together we finally end up with the estimate $$\frac{\mathcal L\{y\}}{\mathcal L\{y'\}}(\frac{2}{h}(\frac{1-z^{-1}}{1+z^{-1}}))\approx \frac{\mathcal Z\{y\}(z)}{\mathcal Z\{y'\}(z)}$$

Question 1. How do we get from the above estimate to the desired estimate

$$\mathcal L\{y\}(\frac{2}{h}(\frac{1-z^{-1}}{1+z^{-1}}))\approx \mathcal Z\{y\}(z)$$

Question 2. It is possible to use trapezoid rule on any linear ODE directly (iterating the rule for higher order derivatives as needed). However, interestingly this gives a different answer than the bilinear transform. The resulting coefficients $a_i$ are the same, however the $b_i$s are different.

So it appears that the bilinear transform distorts the start of the impulse. For example, for 1st order Butterworth lowpass filter we have from bilinear approximation that $b_0=b_1$. As such the impulse response ramps up before ramping down, while the continuous response and trapezoidal estimate only ramps down (see example below). Why are the $b_i$s given by the bilinear transform the "best" (e.g. most widely used) estimates (note: we only consider $b_i$s up to the order of the filter)?

Example: Consider the 1-pole Butterworth filter given by the transfer function:


The differential eq. is given by: $$y=-y'/\omega$$

By trapezoid method and the above eq.: $$y(t+h)\approx y(t)+\frac{h}{2}(y'(t+h)+y'(t))=y(t)-\frac{h\omega}{2}(y(t+h)+y(t))$$ Collecting the terms we get the difference equation $$y[n+1]=\frac{(1-\frac{h\omega}{2})}{(1+\frac{h\omega}{2})}y[n]$$

Which is the same as with bilinear transform, but lacking the term $b_1$.


2 Answers 2


Let me rephrase your derivation of the bilinear transform to clarify the result. For any $t_0<t-T$ we have


Approximating the right-most integral in $(1)$ by the trapezoidal rule we get

$$y(t)\approx y(t-T)+\frac{T}{2}\big(x(t)+x(t-T)\big)\tag{2}$$

Switching over to samples of $y(t)$ and $x(t)$ and using the sloppy but common and handy notation $y[n]=y(nT)$, we define from $(2)$ the discrete-time (DT) system:


This system is now used as the discrete-time approximation to an integrator. In the $\mathcal{Z}$-transform domain it is described by


So the bilinear transform replaces the transfer function of a continuous-time (CT) integrator $H_c(s)=1/s$ by the transfer function $(4)$, i.e., the bilinear transform uses the mapping

$$s\leftrightarrow \frac{2}{T}\frac{z-1}{z+1}\tag{5}$$

So far for part $1$ of your question.

I must admit that I'm not sure I understand part $2$ correctly. Could you give a simple concrete example showing how you obtain two different DT systems from a given CT system?

As for you final question, the bilinear transform is of course not the best way to transform a CT system to a DT system, simply because there cannot be one best way to do so. However, in some applications - such as optimal filter design - it is considered the most useful transform because it preserves optimality of the magnitude. This is the case because the frequency response of the resulting DT system is just a compressed version of the frequency response of the original CT system, i.e., all properties concerning the magnitude (e.g., ripple size, number of zeros in the stopband, etc.) remain unchanged. Just the frequency axis gets warped, but we can take this warping into account when designing the CT system such that the resulting DT system implements the desired band edges.

Furthermore, the bilinear transform is the only direct mapping from the $s$-plane to the $z$-plane that preserves the frequency response in the sense that the $j\omega$-axis is mapped to the unit circle of the $z$-plane. This is not the case with the forward and backward Euler methods. Furthermore, the left half-plane of the complex $s$-plane is mapped to the inside of the unit circle of the $z$-plane, i.e., stability is preserved in the most complete way. This is not the case with the forward Euler method, and even though stability is also preserved with the backward Euler method, the backward Euler method maps the left half-plane to a small region (a circle centered at $z=\frac12$ with radius $\frac12$) inside the unit circle. This means among other things that certain frequency characteristics cannot be achieved with the backward Euler method.

EDIT: Concerning your example of a first order Butterworth filter


you can't ignore the input of the filter in the differential equation, so you get from $(6)$


Now if you use $(7)$ in the trapezoidal approximation

$$y(t)\approx y(t-T)+\frac{T}{2}\big(y'(t)+y'(t-T)\big)\tag{8}$$

you obtain the same difference equation as if you simply had replaced $s$ in $(6)$ by $\frac{2}{T}\frac{z-1}{z+1}$. This must be the case because the bilinear transform simply is the trapezoidal rule (also called Tustin's method).

  • $\begingroup$ Thank you for the answer. Now there is an example of how to obtain a different estimate by trapezoid method. Do you have a reference for bilinear transform preserving the optimality of the magnitude response? $\endgroup$
    – Dole
    Commented Dec 21, 2019 at 20:00
  • 1
    $\begingroup$ @Dole: I've edited my answer to address your example. As for the optimality of the magnitude, since the bilinear transform maps the frequency axis to the unit circle, the frequency response remains unchanged, apart from a frequency warping, so obviously optimal magnitudes remain optimal after the transformation. Ripple size etc. cannot change. $\endgroup$
    – Matt L.
    Commented Dec 21, 2019 at 20:45
  • $\begingroup$ @MattL- Nice answer. Would the Matched-Z qualify if we are not concerned with the aliasing? (Regarding the points in your sentence about the bilinear transform being the only direct mapping). It maps to the jw -axis and maps the left half plane to the complete inner circle (albeit periodically). So for a frequency response that is negligible above half the sampling rate, would the matched Z transform have the same qualities you detail such as optimality of magnitude? $\endgroup$ Commented Dec 22, 2019 at 1:55
  • $\begingroup$ @DanBoschen: My point was to make distinction between what I called "direct mappings", i.e., where $s$ is replaced by a bijective function of $z$ (i.e., mapping in the sense of mapping between complex variables, Moebius transformation), and other transformations, as the impulse invariant or matched z transform. $\endgroup$
    – Matt L.
    Commented Dec 22, 2019 at 11:06
  • 1
    $\begingroup$ @DanBoschen: That's a good question, and the answer is yes and no :) The thing is that aliasing is inherent in the matched z transform, because it can only be applied to a CT rational transfer function (we need poles and zeros after all). And a rational transfer function can never be ideally band-limited, so aliasing is always there by definition. But we can get close to reproducing the frequency response exactly if the sampling rate is sufficiently high, and if the CT system has a low pass characteristic. $\endgroup$
    – Matt L.
    Commented Dec 22, 2019 at 13:06

Here's another "justification" for the Bilinear Transform. If you equate the transfer functions, in the $s$ and $z$ domains of the unit sample delay, you have:

$$ z^{-1} = e^{-sT} $$

which says the same thing as

$$ z = e^{sT} $$

where $T$ is the sampling period or the reciprocal of the sample rate:

$$ T \triangleq \frac{1}{f_\mathrm{s}} $$

So the bilinear approximation (which is equivalent to the trapezoidal rule or Tustin's method) is made to that exponential.

$$\begin{align} z &= e^{sT} \\ \\ &= \frac{e^{sT/2}}{e^{-sT/2}} \\ \\ &\approx \frac{1+sT/2}{1-sT/2} \\ \\ \end{align}$$

If you solve for $s$, you get:

$$ s \leftarrow \frac{2}{T} \, \frac{z-1}{z+1} = \frac{2}{T} \, \frac{1-z^{-1}}{1+z^{-1}} $$

Now this approximation is normally good for small $s$ but it turns out that every point on the $j \omega$-axis is mapped to a point on the unit circle of the $z$-plane


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.