Isabel Kaspriskie

In this fourth and final installment of the Demystifying Digital Filters series, we’ll take our first steps into the implementation of IIR filter designs. While the field of recursive digital filters is enormous, understanding the basic landscape will help you to navigate and take your first steps into exploring these new territories.

When we design a filter, we’re usually referring to optimizing for a set of “design” parameters such as the steepness of the transition band or the flatness of a stopband or passband. We do this by describing the filter mathematically and then solving a set of equations to get the corresponding filter coefficients.

Since analog filter design has been studied for decades, many of the most useful design equations have already been solved. These solutions are known as “filter prototypes.” 

The utility of a prototype filter comes from the property that all these other filters can be derived from it by applying a scaling factor to the components of the prototype. The filter design need thus only be carried out once in full, with other filters being obtained by simply applying a scaling factor.

There are a several techniques that allow us to use these prototypes in the digital domain, including:

Before we discuss these digitizing techniques though, let's look at some of the most common filter prototypes — Butterworth, Chebyshev, and elliptic.

pasted image

Keep in mind that the y-axis of these frequency response plots are in units of dB (on a logarithmic scale) so the large “bumps” are less audible than they appear. All filters represented here share the same order and cutoff value (and passband/stopband ripple allowances, as appropriate).

 is named after Stephen Butterworth, an electrical engineer who first 

 the design in 1930. The goal of the Butterworth filter is to achieve a 

 — there should be no distortion or “ripples.” Out of the four prototypes here, Butterworth filters have the slowest rolloff for a given filter order.

 used to derive this filter design. Both Type-I and Type-II Chebyshev filters aim to minimize the error between an ideal and actual filter across the frequency response.

The rolloff of both types of filters are steeper than that of Butterworth filters — they do this by allowing ripple in the passband (for Type-I) or ripple in the stopband (for Type-II).

Type-I Chebyshev filters have slightly faster rolloffs than Type-II, but the ripple in the passband (which is likely low enough to be inaudible) may be less acceptable than ripple in the stopband.

 (also known as Cauer or Zolotarev filters) have an even steeper rolloff than either of the Chebyshev filters. They achieve this by allowing ripple in both the stopband and passband. The amount of ripple in either band can be adjusted independently from the other.

Before we begin discussing digitization methods, let’s start by reminding ourselves of the distinction between analog and digital systems and signals.

In an analog signal, there is no limit to how far you can “zoom in” to find a value at any point; for any time 

 along the signal, there’s an exact value available. This is referred to as having a 

The time domain of a digital signal, on the other hand, is 

. This means that a finite number of samples 

 along the signal are defined. If you want to know the value at a point that isn’t defined, you have to use the surrounding samples and 

 to make a reasonable guess about what that value is.

Since analog filter design equations are already available thanks to the work of electrical engineers, all we need to do is to be able to change the domain for these equations from the continuous analog representation to the discrete digital domain. If that sounds scary, just remember that you’re actually already a pro at changing domains! Any time in the course of this tutorial series that you’ve worked with a “transform” such as a Fourier transform or a Z-transform, you’ve been moving between domains. All that those transformations do is change the dependent variable of an equation by providing some kind of map between them.

For instance, you’re probably already familiar with how to transform continuous-time signals to discrete-time ones. If you have a sine wave signal 

, in the continuous-time domain (meaning 

image.png

If you wanted that same signal to be transformed to the discrete-time domain (where samples are defined for integer values 

), you’ll need to have some kind of relationship between 

. The easiest way to do that is to define a constant 

 as the distance between samples (the sampling interval), and get the following relationship between 

By substituting that relationship into the continuous-time signal representation, you would get back out a discrete-time representation that only depends on integer values 

Okay, so if the relationship between continuous- and discrete-time signals is so simple, could it be just as easy to transform between continuous and discrete systems?

Here’s the bad news first: notice that when you go from analog to digital signal representations, you’re losing information. All of the data between samples in a digital signal are lost! The only way to find out the values between data points is to make some kind of educated guess as to what the value would be (using interpolation). The same goes for mapping the continuous-time equations governing analog systems to equations for digital ones. By making any kind of mapping, you will lose some fidelity with digitization.

The good news is that with digitizing analog filter prototypes, you can choose 

 of fidelity you are willing to give up in exchange for improvements in other areas. Let’s explore the characteristics of three different transformations from the analog to the digital domain.

 takes a continuous-time impulse response 

 from an analog filter, samples it in intervals of 

A characteristic of the impulse invariant transformation is an 

 frequency response due to the spectral folding explained by the sampling theorem.

 is mapping between the analog and digital transfer functions.

. By using the following relationship between 

, you can take an analog transfer function (in terms of 

Written another way, the relationship between the analog and digital transfer functions is:

A characteristic of the bilinear transform is maintenance of stability. Since the bilinear transform maps the 

 to the z-plane, stable poles of an analog filter will always maintain their stable position when mapped to a different domain. One downside of the bilinear transform is “frequency warping” — the frequency response of a digitized filter is 

 (also known as the root matching method) works by taking all poles and zeros 

 of the analog filter and mapping them to the Z-plane with the following relationship:

This method is simple and maintains stability of the original analog filter, but it warps both the time- and frequency-based characteristics of the digital filter result. Since that’s the case, both the impulse invariant and bilinear transforms are usually the transforms of choice for filter design.

As you build higher and higher order filters, you will notice that the order of magnitude of the coefficients start to grow further apart. For example, a sixth-order filter can have some coefficients with a magnitude of 10^(-10) and others with a magnitude of 10^(-1). These differences in order of magnitude can lead to major floating point errors and numerical instability.

To minimize this kind of numerical error, higher order filters can be split up into smaller, lower order “elementary filter sections” (a term coined by your good friend Julius O. Smith III) like one-pole, one-zero, two-pole, and two-zero filters.

This “splitting” of higher order filters is possible due to two properties of transfer functions:

1. Transfer functions in series can be combined by multiplying

 into two separate systems where the first section outputs 

The transfer function of the “lumped” system is:

You can also separately write the transfer functions for each section in the series version.

Let’s check that multiplying those transfer functions will give us the same result as the lumped system.

Great! This property of transfer functions is true.

2. Transfer functions in parallel can be combined by adding

How about the other case of sections in parallel?

Perfect — summing sections in parallel works as expected.

To summarize, if you can turn a high order transfer function into a product of lower order transfer functions, you can take advantage of these properties and implement the higher order filter as a set of lower order filters in series. A similar idea would work for creating a sum of lower order transfer functions and implementing them as a set of parallel filters.

The biquadratic (“biquad”) filter is a two-pole, two-zero filter that is the workhorse building block of IIR filters. Since it is a second order filter, it provides numerical stability. Also, the digital implementation is flexible — you can easily make any of the other elementary filter sections (like a one-pole or one-zero filter) by simply using zero as a filter coefficient.

There are four different forms of the biquad filter:

The details of these different forms are out of the scope of this series, but it is worthwhile to know they exist since you may see different variations of the biquad out in different block diagrams. Different forms have different strengths and weaknesses, and you can learn more about them in 

Introduction to Digital Filters with Audio Applications

In practice, when building higher-order IIR filters, series implementations are generally preferred. This technique is known as “cascading.”

Armed with an understanding of how IIR filters are designed, you can explore the 

 objects in Max with a better idea of what’s occurring behind the scenes. The 

 package (available from the Max Package manager) also contains a wide variety of IIR filters for both MSP and Jitter applications.

Now that you’ve made it to the end of this whirlwind tour of digital filters, you should have new insight into what a digital filter is, how one is implemented, and what design decisions can be made.

If you would like an all-in-one PDF copy of this series, check out the download below. Print it out, put it in a conspicuous place, and impress all of your friends with your new knowledge!

demystifying-digital-filters.pdf

Since analog filter design has been studied for decades, many of the most useful design equations have already been solved. These solutions are known as “filter prototypes.” Wikipedia nicely summarizes their usefulness:

The Butterworth filter is named after Stephen Butterworth, an electrical engineer who first published the design in 1930. The goal of the Butterworth filter is to achieve a maximally-flat passband — there should be no distortion or “ripples.” Out of the four prototypes here, Butterworth filters have the slowest rolloff for a given filter order.

Chebyshev filters are named for the Chebyshev polynomials used to derive this filter design. Both Type-I and Type-II Chebyshev filters aim to minimize the error between an ideal and actual filter across the frequency response. 

The rolloff of both types of filters are steeper than that of Butterworth filters — they do this by allowing ripple in the passband (for Type-I) or ripple in the stopband (for Type-II). 

Elliptic filters (also known as Cauer or Zolotarev filters) have an even steeper rolloff than either of the Chebyshev filters. They achieve this by allowing ripple in both the stopband and passband. The amount of ripple in either band can be adjusted independently from the other. 

In an analog signal, there is no limit to how far you can “zoom in” to find a value at any point; for any time t along the signal, there’s an exact value available. This is referred to as having a continuous-time domain.

The time domain of a digital signal, on the other hand, is discrete. This means that a finite number of samples n along the signal are defined. If you want to know the value at a point that isn’t defined, you have to use the surrounding samples and interpolate to make a reasonable guess about what that value is.

For instance, you’re probably already familiar with how to transform continuous-time signals to discrete-time ones. If you have a sine wave signal f(t), in the continuous-time domain (meaning t can be any value), you would write:

If you wanted that same signal to be transformed to the discrete-time domain (where samples are defined for integer values n), you’ll need to have some kind of relationship between t and n. The easiest way to do that is to define a constant T as the distance between samples (the sampling interval), and get the following relationship between t and n:

By substituting that relationship into the continuous-time signal representation, you would get back out a discrete-time representation that only depends on integer values n. 

The good news is that with digitizing analog filter prototypes, you can choose which kinds of fidelity you are willing to give up in exchange for improvements in other areas. Let’s explore the characteristics of three different transformations from the analog to the digital domain.

The impulse invariant transformation takes a continuous-time impulse response h_c(t) from an analog filter, samples it in intervals of T, and uses that impulse h[n] as the basis for the digital filter.

A characteristic of the impulse invariant transformation is an aliased frequency response due to the spectral folding explained by the sampling theorem. 

The bilinear transform is mapping between the analog and digital transfer functions.

In continuous-time systems, instead of z, transfer functions use the variable s. By using the following relationship between s and z, you can take an analog transfer function (in terms of s) and digitize it.

A characteristic of the bilinear transform is maintenance of stability. Since the bilinear transform maps the s-plane to the z-plane, stable poles of an analog filter will always maintain their stable position when mapped to a different domain. One downside of the bilinear transform is “frequency warping” — the frequency response of a digitized filter is slightly different from its analog counterpart.

The matched Z-transform (also known as the root matching method) works by taking all poles and zeros s of the analog filter and mapping them to the Z-plane with the following relationship:

As you build higher and higher order filters, you will notice that the order of magnitude of the coefficients start to grow further apart. For example, a sixth-order filter can have some coefficients with a magnitude of 10^(-10) and others with a magnitude of 10^(-1). These differences in order of magnitude can lead to major floating point errors and numerical instability. 

Imagine you are splitting up H(z) into two separate systems where the first section outputs v(n) as the input to the second section.

Great! This property of transfer functions is true.

2. Transfer functions in parallel can be combined by adding

Let’s check that summing H_1 and H_2 returns the expected result.

From the diagram, you can see that y_1 and y_2 sum to y, so you can now write:

To summarize, if you can turn a high order transfer function into a product of lower order transfer functions, you can take advantage of these properties and implement the higher order filter as a set of lower order filters in series. A similar idea would work for creating a sum of lower order transfer functions and implementing them as a set of parallel filters. 

The details of these different forms are out of the scope of this series, but it is worthwhile to know they exist since you may see different variations of the biquad out in different block diagrams. Different forms have different strengths and weaknesses, and you can learn more about them in this section of Introduction to Digital Filters with Audio Applications.

In practice, when building higher-order IIR filters, series implementations are generally preferred. This technique is known as “cascading.” 

Armed with an understanding of how IIR filters are designed, you can explore the biquad~, cascade~, filterdesign, and filtergraph~ objects in Max with a better idea of what’s occurring behind the scenes. The Filter package (available from the Max Package manager) also contains a wide variety of IIR filters for both MSP and Jitter applications.

filter_types.jpg

In this final installment, we’re going to take our first steps into the design and implementation of IIR filters. While the field of recursive digital filters is enormous, you’ve now got the basic tools to take your first steps into exploring these new territories.

demystifying-digital-filters-part-4

So great, thank you again! And all-in-one .pdf, too 🙌

https://cycling74.com/forums/announcing-a-new-package-filter

So great, thank you again! And all-in-one .pdf, too 🙌
and i see what you did there 😉:
https://cycling74.com/forums/announcing-a-new-package-filter

Thanks again for all of it! 🍻

yes lots of thanks !!! although i didn't understand everything, this series did help me to see more clearly in the mystical world of filters :) also thanks for the nice package indeed !

Thanks a lot for your descriptions, I like the graphical dragging of poles and zeros for the biquad filter.

Connected to a saw generator student can hear and watch building a filter of their own. Great!

Thanks a lot for your descriptions, I like the graphical dragging of poles and zeros for the biquad filter.
Connected to a saw generator student can hear and watch building a filter of their own. Great!

Is it worth it getting deeper into filters and making your own? What can you hope to achieve that isn't possible with "standard" filters like FabFilter etc.

 you want. that is quite a huge difference from when it is not by yourself / not available as max patch.

try to modulate all 17 channels of a fabfilter VST plug-in with an audiosignal and you will see.


if you build your own you can do anything you want. that is quite a huge difference from when it is not by yourself / not available as max patch.

try to modulate all 17 channels of a fabfilter VST plug-in with an audiosignal and you will see.


That's a reasonable question, and not silly at all.

Commercial filters offer a lot of amazingly feature-rich options. In particular, they're usually well optimized for computational efficiency and ease of use. They take the mathematically-optimized designs that minimize things like distortion of frequencies and phases. If you're mixing a song with dozens of tracks on it, you're probably going to want a filter that's been optimized in these ways and has been designed for UI and UX.

But what if you actually want distortions? Maybe you want really specific kinds analog modeling that just aren't available on a computer yet. If you have learned about digital filters and you happen to find a circuit diagram of that analog filter, you could find the transfer function from that circuit diagram and transform to the digital domain. Then you don't have to wait for someone else to model the filter for you.

That, I think, points to the biggest benefit of knowing about how filters (and DSP in general) work. If you know how things work, it doesn't matter what environment you're working in, whether it's 

, C++, or some other language. You have the deep knowledge that frees you from needing to rely on someone else to have written a filter function/object/library for you. You can optimize the things 

Also, the DSP concepts you can learn by studying filters are a foundation for many other kinds of digital audio effects, from phasers, to flangers, to reverbs.

But what if you actually want distortions? Maybe you want really specific kinds analog modeling that just aren't available on a computer yet. If you have learned about digital filters and you happen to find a circuit diagram of that analog filter, you could find the transfer function from that circuit diagram and transform to the digital domain. Then you don't have to wait for someone else to model the filter for you. 

That, I think, points to the biggest benefit of knowing about how filters (and DSP in general) work. If you know how things work, it doesn't matter what environment you're working in, whether it's gen~, C++, or some other language. You have the deep knowledge that frees you from needing to rely on someone else to have written a filter function/object/library for you. You can optimize the things you care about. That's the biggest win.

This discussion is really helpful - reading it raised some questions about IIR filter prototypes.

With Chebyshev, Elliptic, etc. ripple can exit in the passband or stopband segments. 

When does ripple become audible and how does this impact sound output - like with noise or simple oscillators as input? 

Normally I think we want to minimize ripple - but can we use ripple creatively - for experimental sound filters, creative processing? 

it is probably correct to say that there are multiple methods you could use.

for example if you look into what filterdesign calls "chebycheff II" - can you imagine what happens when you upsample that x2? or simply hook up a butterworth lowpass at 17khz?

moving the ripples at either end of the spectrum is already a great performance of this topology, which you can make use of to "remove" them.

Demystifying Digital Filters, Part 4