IIR Filters can be implemented in many ways

• As a recursive filter with numerous possible architectures

• It can be challenging to assure stability given arithmetic accuracy
requirements.

• As a series of cascaded Biquadratic filters

• This is a common approach
• Can yield varying computational accuracy requirements depending on how
poles and zeros are combined in the Biquad sections
 Brief history of development of PM Algorithm

The students [at Rice University] in the "Signal Theory" course [in 1971]
were required to do a project and since Chebyshev approximation was a
major topic in the course, the implementation of this new algorithm
became James McClellan's course project. This ultimately led to the
Parks–McClellan algorithm, which involved the theory of optimal
Chebyshev approximation and an efficient implementation. By the end of
the spring semester, McClellan and Parks were attempting to write a
variation of the Remez exchange algorithm for FIR filters. It took about six
weeks to develop and some optimal filters had been designed
successfully by the end of May.
 Important insights

• These filter designs are not the result of solving closed-form

equations.
• They are the result of development of efficient algorithms
(e.g. Remez Exchange) to find an iterative solution.
• Different approaches will lead to different results, sometimes
with very similar performance.
• There are tradeoffs related to cost and performance.
• Like many areas of technology, there is not a single “right”
answer.
 Alternate form of the FIR filter

𝑥𝑛
𝑝3 𝑝2 𝑝1 𝑝0

+ + + + 𝑦𝑛
+ Z-1 + Z-1 + Z-1 +
+ + + +
 Alternate Form adapted to IIR Filter

𝑥𝑛
𝑝3 𝑝2 𝑝1 𝑝0

+ + + + 𝑦𝑛
+ Z-1 + Z-1 + Z-1 +
+ + + +

−𝑑3 −𝑑2 −𝑑1

Another Possible IIR Architecture
 The Biquadratic Filter

A second-order recursive linear filter

Its Transfer function is the ratio of two quadratic functions

Higher-order IIR filters may be implemented as serially-connected

biquad sections to reduce risk of instability
 The Biquadratic (Biquad) Filter
𝑌 𝑧 𝑏0 + 𝑏1 𝑧 −1 + 𝑏2 𝑧 −2
𝐻 𝑧 = =
𝑋 𝑧 𝑎0 + 𝑎1 𝑧 −1 + 𝑎2 𝑧 −2

𝑌 𝑧 𝑎0 + 𝑎1 𝑧 −1 + 𝑎2 𝑧 −2 = 𝑋 𝑧 𝑏0 + 𝑏1 𝑧 −1 + 𝑏2 𝑧 −2
1
𝑌 𝑧 = 𝑋 𝑧 𝑏0 + 𝑏1 𝑧 −1 + 𝑏2 𝑧 −2 − 𝑌 𝑧 𝑎1 𝑧 −1 + 𝑎2 𝑧 −2
𝑎0

1
𝑦 𝑛 = 𝑏0 𝑥 𝑛 + 𝑏1 𝑥 𝑛 − 1 + 𝑏2 𝑥 𝑛 − 2 − 𝑎1 𝑦 𝑛 − 1 − 𝑎2 𝑦 𝑛 − 2
𝑎0
Normalized (divide all coefficients by 𝑎0 )

𝑦 𝑛 = 𝑏0 𝑥 𝑛 + 𝑏1 𝑥 𝑛 − 1 + 𝑏2 𝑥 𝑛 − 2 − 𝑎1 𝑦 𝑛 − 1 − 𝑎2 𝑦 𝑛 − 2
BiQuad Direct Form I Implementation
𝑦 𝑛 = 𝑏0 𝑥 𝑛 + 𝑏1 𝑥 𝑛 − 1 + 𝑏2 𝑥 𝑛 − 2 − 𝑎1 𝑦 𝑛 − 1 − 𝑎2 𝑦 𝑛 − 2

https://en.wikipedia.org/wiki/Digital_biquad_filter
 Biquad Direct Form II Implementation
𝑦 𝑛 = 𝑏0 𝑥 𝑛 + 𝑏1 𝑥 𝑛 − 1 + 𝑏2 𝑥 𝑛 − 2 − 𝑎1 𝑦 𝑛 − 1 − 𝑎2 𝑦 𝑛 − 2

https://en.wikipedia.org/wiki/Digital_biquad_filter
Some considerations with integer math
• Multiplying a sample of n bits by a coefficient of m bits results in a
product of length m + n bits.
• Adding two of these m + n bit products can result in a sum that is
m + n + 1 bits long.
• With the chain of adders we need to support m + n + 3 bits.
• With the multiplication of the 𝑦 𝑛 values the results will continue to
grow each time interval unless truncation (rounding) happens with
fixed-point implementations.
• With floating-point implementations, rounding is automatic based on
the precision of the floating-point values
Transposition of Direct Form I

https://en.wikipedia.org/wiki/Digital_biquad_filter
Transposition of Direct Form II

https://en.wikipedia.org/wiki/Digital_biquad_filter
Fixed vs Floating Point Implementations
Fixed point DSP usually prefers the non transposed forms and
has an accumulator with a large number of bits and is
rounded when stored in main memory. Floating point DSP
usually prefers the transposed form, each multiplication and
potentially each addition are rounded; the additions are
higher precision result, when both operands have similar
magnitude.

https://en.wikipedia.org/wiki/Digital_biquad_filter
Add Quantizer Q(z) to Direct Form I

Truncation to
reduce bit length

This adds
quantization noise

https://en.wikipedia.org/wiki/Digital_biquad_filter
Adding Quantization to the Direct Form II

https://en.wikipedia.org/wiki/Digital_biquad_filter
 Cascading Biquads

𝑦0 𝑛 𝑤1 𝑛 𝑦1 𝑛 𝑤2 𝑛 𝑦2 𝑛 𝑤3 𝑛 𝑦3 𝑛
𝑏01 𝑏02 𝑏03
𝑥𝑛 y𝑛
𝑧 −1 𝑧 −1 𝑧 −1
𝑎11 𝑏11 𝑎12 𝑏12 𝑎13 𝑏13

𝑧 −1 𝑧 −1 𝑧 −1
𝑎21 𝑏21 𝑎22 𝑏22 𝑎23 𝑏23

𝑁𝑠
𝑏0𝑘 + 𝑏1𝑘 𝑧 −1 + 𝑏2𝑘 𝑧 −2
𝐻 𝑧 =ෑ
1 − 𝑎1𝑘 𝑧 −1 − 𝑎2𝑘 𝑧 −2
𝑘=1
 𝑦0 𝑛 𝑤1 𝑛 𝑦1 𝑛 𝑤2 𝑛 𝑦2 𝑛 𝑤3 𝑛 𝑦3 𝑛
𝑏01 𝑏02 𝑏03
𝑥𝑛 y𝑛
𝑧 −1 𝑧 −1 𝑧 −1
𝑎11 𝑏11 𝑎12 𝑏12 𝑎13 𝑏13

𝑧 −1 𝑧 −1 𝑧 −1
𝑎21 𝑏21 𝑎22 𝑏22 𝑎23 𝑏23

𝑦0 𝑛 = 𝑥 𝑛

𝑤𝑘 𝑛 = 𝑎1𝑘 𝑤𝑘 𝑛 − 1 + 𝑎2𝑘 𝑤𝑘 𝑛 − 2 + 𝑦𝑘−1 𝑛 , 𝑘 = 1,2, ⋯ , 𝑁𝑠

𝑦𝑘 𝑛 = 𝑏0𝑘 𝑤𝑘 𝑛 + 𝑏1𝑘 𝑤𝑘 𝑛 − 1 + 𝑏2𝑘 𝑤𝑘 𝑛 − 2 , 𝑘 = 1,2, ⋯ , 𝑁𝑠

𝑦 𝑛 = 𝑦𝑁𝑠 𝑛
Combining Pairs
• There are many ways of pairing poles and zeros in these second order
sections.
• In theory all combinations will produce the same result when infinite-
precision arithmetic is used.
• Given the limitations of finite-precision arithmetic, the results may be
quite different with different pairings.