Bilinear Transformation Made Easy

Peter J. Pupalaikis Senior Product Development Engineer LeCroy Corporation 700 Chestnut Ridge Road, Chestnut Ridge NY 10977 (914) 425-2000 x 3174 [email protected]

Abstract
A formula is derived and demonstrated that is capable of directly generating digital filter coefficients from an analog filter prototype using the bilinear transformation. This formula obviates the need for any algebraic manipulation of the analog prototype filter and is ideal for use in embedded systems that must take in any general analog filter specification and dynamically generate digital filter coefficients directly usable in difference equations. the formula which is the subject of this paper, all of the steps in this method are identical to traditional methods of design: [1] 1. Pre-warp the filter frequencies of interest. This involves replacing the critical frequencies in the analog design criteria f c with the following:

Fc =
2.

1. Introduction
Digital filters are often designed by transforming analog filter designs. One of the most popular transformation methods is the bilinear transformation. The bilinear transformation is a simple algebraic substitution in which s, the integration operation in the s domain, is replaced with the equivalent trapezoidal integration function in the z domain:

fs f c tan fs
(2)

Design the analog filter according to the design criteria using the pre-warped frequencies, generating the analog s domain transfer function of the following form:

H (s ) = n = 0
N

an s n bn s n
(3)

s 2 fs

1 z 1 1+ z
1

n =0

(1)

f s is the digital sampling frequency in Hz.

In order to realize the transfer function as a difference equation, the equation must be simplified and all terms collected such that the transfer function represents a ratio of two polynomials in z . This work is easily carried out using mathematics packages capable of symbolic manipulation of equations, but this capability is not generally present in an embedded digital system. In the past, this has precluded the use of the bilinear transformation in systems that generate digital filters dynamically from analog prototypes.
1

N is the maximum order of the numerator and denominator polynomials. In the case where the two polynomials are not of the same order, set the non-existent coefficients to zero. 3. Generate the digital filter transfer function and difference equation of the following forms: Transfer Function:

H (z ) = n = 0
N n=0

An z n Bn z n
(4)

2. Filter Design Method Using The Bilinear Coefficient Formula

This paper presents a manner in which the simplified z domain transfer function is generated with the equation immediately in the proper form needed for direct implementation of the digital filter as a difference equation. Note that except for the use of

Difference Equation: N A N B y[k ] = x[k n] n y[k n] n B0 B0 n =1 n=0

(5)

The digital filter coefficients An and Bn in (4) and (5) are calculated using the following equations:

3. C Code Implementation
This small piece of C code is provided which generates filter coefficients using (6), (7), and (8): //------------------------------------------------------// Digital Filter Coefficient Generator using // the Bilinear Coefficient Formula. // Peter J. Pupalaikis //------------------------------------------------------// simple factorial function static double f(double i) { if (i==0) return 1; else return i*f(i-1); } #define max(a,b) (((a)>(b))?(a):(b)) #define min(a,b) ((a<b)?a:b) // computes a digital filter coefficient from the analog // coefficients using the bilinear coefficient formula double s_to_z ( double a[], // s domain coefficients int N, // index of last coefficient (order) int n, // z domain coeffient desired double fs // sample rate ) { double z_coef = 0.0; for (int i=0; i<= N; ++i) { double acc = 0.0;

An =
and

a f
i i =0

i s

BF (i, n, N )
(6)

Bn =
where: BF (i, n, N ) =

b f
i i =0

i s

BF (i, n, N )
(7)

2i

i!(N i )!( 1)k k!(i k )!(n k )!(N i n + k )! k = max(0, n N + i )

min(n, i )

N i0 , N n0
(8)

I call (6) and (7) the bilinear coefficient formula and (8) a bilinear factor. These equations are remarkably simple and offer a programmatic way of filter coefficient generation. Here is an abbreviated table of Bilinear Factors:
N 0 1 2 i 0 0 1 0 1 2 0 1 2 3 0 1 2 3 4 0 1 2 3 4 5 0 1 2 3 4 5 6 0 1 1 2 1 2 4 1 2 4 8 1 2 4 8 16 1 2 4 8 16 32 1 2 4 8 16 32 64 1 1 -2 2 0 -8 3 2 -4 -24 4 4 0 -16 -64 5 6 4 -8 -48 -160 6 8 8 0 -32 -128 -384 2 n 3 4 5 6

1 -2 4 3 -2 -4 24 6 0 -8 0 96 10 4 -8 -16 32 320 15 10 -4 -24 -16 160 960

1 -2 4 -8 4 -4 0 16 -64 10 -4 -8 16 32 -320 20 0 -16 0 64 0 -1280

1 -2 4 -8 16 5 -6 4 8 -48 160 15 -10 -4 24 -16 -160 960

1 -2 4 -8 16 -32 6 1 -8 -2 8 4 0 -8 -32 16 128 -32 -384 64 Table 1

for (int k=max(n-N+i,0); k<=min(i,n); ++k) { acc += ( ( f(i)*f(N-i) ) / ( f(k)*f(i-k)*f(n-k)*f(N-i-n+k) ) ) * ( (k & 1) ? -1.0 : 1.0 ); } z_coef += (a[i] * pow(2.0*fs,i) * acc); } return z_coef; }

4. Derivation
In order to derive the bilinear coefficient formula, start by realizing that any s-domain transfer function can be written in the form as (3), where N is the maximum order of the numerator and denominator polynomials. When the polynomials are not of the same order, zeros are substituted for the non-existent coefficients. The bilinear transformation involves substituting for the Laplace variable s, as specified in (1). When this substitution is performed, (3) becomes:

min(i , N a ) Na Nb Na + Nb i a xi b xi = x ak bi k i i i =0 k = max(0, i N b ) i =0 i =0

(15)

Using (15), (14) can be written as:

n =0

z i an 2 n f s n i =0

n N n k k ik ( 1) k =max( 0,i N + n )
min( i ,n )

(16)

By altering the order of summation, and making a change of variables, (16) becomes:
n =0

H (z ) =

n =0 N n =0

an 2n f s n 2 fs
n n

(1 z (1 + z (1 z (1 + z

1 n 1 n 1 n 1 n

) ) ) )

a 2i f s i z n i i =0

i N i k 1 ( ) k n k k =max(0,n N +i )
min(n,i )

(17)

or:
n =0

a f i 2i z n i s i =0

i!(N i )!( 1)k k!(i k )!(n k )!(N i n + k )! k = max(0,n N +i )

min(n,i )

(9)

(18)

Multiplying

(1 + z )

numerator
N

and

denominator

by

1 N

, (9) becomes:

H (z ) =

a b
n=0 n =0 N

2 n f s n 1 z 1 2 f s 1 z
n n

) (1 + z )
n

1 N n

At this point, it becomes clear that the z domain transfer function can be written as (4), where the filter coefficients An and Bn are those directly calculated by (6), (7), and (8). This concludes the derivation.

1 n

) (1 + z )

1 N n

5. Design Methodology
(10)

At this point, the derivation will deal only with the numerator of (10) with the understanding that the resulting formula will be valid for the denominator, as well, with minor modification. First, observe that 1 + z 1 and 1 z 1 can be expressed using the following identities:

Section 2 - "Filter Design Method Using The Bilinear Coefficient Formula" outlined the steps involved in designing digital filters using analog prototypes and transforming these analog filters to digital filters using the bilinear coefficient formula. Figure 1 shows a flowchart of this process:
Frequency Domain Design Parameters design starting points

(1 + x )m
where:

m i m m i i = i x and (1 x ) = i ( 1) x i =0 i =0
m

(11) and (12)

Prewarp Critical Frequencies Design Analog Filter Analog Filter Coefficients

a a! b = b!(a b )!
(13)

(13) is known as a binomial coefficient. [2] Utilizing (11) and (12), and (13), the numerator of (10) can be expressed as: N n n N n N n i i i ( ) an 2 n f s n 1 z z n =0 i =0 i i =0 i

Substitution and Algebraic Manipulation

Bilinear Coefficient Formula

Difference Equation Implementation Digital Filter Designed

(14)

We can then make use of another identity:

Figure 1

In Figure 1, the process of substitution and the algebraic manipulation of the filter has been crossed out and replaced with the bilinear coefficient formula. The benefit of this formula in embedded systems is readily apparent. The theory behind and design considerations of the use of the bilinear transformation are treated extensively in many engineering textbooks. This paper will not attempt to duplicate this effort - only present the bilinear coefficient formula and some design considerations resulting from the use of this formula. Figure 1 shows two design starting points. Each is considered subsequently: 5.1 Frequency Domain Design Parameters Filter design using frequency domain design parameters is the method of choice for designing Butterworth, Chebyshev, Bessel, and (other) elliptic filters. The method outlined in section 2 is well equipped for handling these designs. Furthermore, there is a large body of knowledge, software, and algorithms that exist for the design of these types of analog filters. The only consideration necessary is that of pre-warping all frequency specifications using (2). This means that the digital filter must be designed by translating the analog filter frequency criteria, such as pass-band and stop-band edge frequencies using (2). Note that filters generated in this manner will be guaranteed to meet the design requirements only if the pre-warped frequencies are considered when determining the order of the filter. 5.2 Analog Filter Coefficients Filter design using analog filter coefficients is obviously the method of choice when the filter specification is the analog filter coefficients themselves. It is also the method of choice when the task of the digital filter is to emulate an analog system, in which case it might be unknown at what frequencies close emulation is important. There is no opportunity for pre-warping of frequency design criteria when nothing is known specifically about the filter coefficients and how they were generated. This is not necessarily a problem if the frequencies of interest are much lower than the sample rate of the system. Considering (2), this becomes apparent. If fc fc f c << f s , then tan and f f s s therefore Fc f c . In this case, the filters designed with and without frequency pre-warping are almost identical anyway.

6. Practical Considerations
It is important to note that the digital filter generated using the coefficient formula is mathematically equivalent to that generated through algebraic substitution and manipulation to the extent that the frequency response and time domain performance are identical. The actual filter structure or topology, however, may not be identical to that generated through other digital filter design techniques. Generally, in the design of digital filters, round-off error and numerical overflow/underflow are addressed by breaking the filter into smaller pieces (usually biquad sections). Analog filters (especially active filters) are sometimes designed in sections, so the transformation of the analog filter sections themselves is possible. For high order analog filters expressed as in (3), keep in mind there may be numerical problems in the direct non-sectioned filter implementation.

7. Simple Example
Consider a 3rd order analog normalized Butterworth low-pass filter. Its transfer function is: 1 H ( s) = 1 + 2 s + 2 s 2 + s3 The poles of this filter are 1 , 0.5 + j 0.866 , and 0.5 j 0.866 . They are all located on a circle at a distance of 1 from the origin. Using a normalized sample rate of f s = 1 Hz, the critical frequency of = 1 translates to a pre-warped frequency of f = 0.174, or = 1.093 . Multiplying each pole by 1.093 correctly pre-warps the filter, and the new pole locations are 1.093 , 0.546 + j 0.946 , and 0.546 j 0.946 . These pole locations are shown in Figure 2:
.
1

(19)

1 1 0 1

analog filter poles prewarped pole locations

Figure 2

The transfer function of the analog filter with these new pole locations is: 1.3057 H ( s) = 1.3057 + 2.3892 s + 2.1860 s 2 + s 3
(20)

The analog filter is plotted at s = j 2 f while the digital filter is plotted at z = e

j 2 f fs

8. Conclusion
The bilinear coefficient formula was presented, derived and demonstrated. This formula was shown to be a powerful tool for filter generation. Once the design considerations are understood, there is a real benefit to be gained by using these formulae in systems with all or some of the following criteria: 1. 2. A requirement for dynamic filter generation with flexible filter design criteria. A requirement for varying order filters (usually filters whose order depends on the performance needed). [1] A requirement for IIR filters (or more specifically, a lack of requirement for FIR filters). Usually this means that linear phase is not a necessity. Sometimes, this means that the digital filter is needed to emulate, but not necessarily outperform an analog filter with regard to phase response. If analog filter coefficients are being used directly, the system must be significantly oversampled (i.e. the sample rate must be much higher than the frequencies of interest). The system must be able to deal with the potential numerical problems resulting from high order, non-sectioned filters - usually through the use of high precision/range floating point arithmetic and number format.

The filter has been adjusted to maintain unity gain at 0 Hz. So far in this example, nothing is germane to the usage of the bilinear coefficient formula. The work done in pre-warping the filter was done only so that the results will match those found in filter design textbooks. Since the transfer function is in the form of (3), the bilinear coefficient formula in (6) and (7) is applied, and the resulting coefficients are: 22.828 1.306 3 . 917 24.048 A= and B = 14.395 3.917 2.729 1.306 It is common to divide all of the numerator coefficients by A0 and the denominator coefficients by B0 and to generate a denominator scaling term

3.

4.

D0 =

B0 . Doing so results in: A0

5.

1 1 3 1.053 , and D0 = 17.484 A= ,B= 0.631 3 0.12 1 These results compare favorably with those obtained from tables of normalized Butterworth low-pass digital filter coefficients: [3]
Figure 3 shows the magnitude response plotted as 20 Log ( H ) where H is given by (19) for the analog filter and (4) for the digital filter.
.
10

These equations facilitate the use of dynamic digital filter design using bilinear transform techniques in embedded systems.

9. References
[1] Parks, T.W. and Burrus, C.S. Digital Filter Design, John Wiley and Sons, Inc., 1987 [2] Knuth, Donald E., The Art of Computer Programming Volume I, Addison-Wesley, 1997 [3] Jong, M. T. Methods of Discrete Signal and System Analysis, McGraw-Hill, 1982.

10

20 0.01

0.1

analog prototype digital filter designed using bilinear coefficient formula digital filter obtained from textbook

Figure 3