# Transfer Function Analysis

This chapter discusses filter*transfer functions*and associated analysis. The transfer function provides an algebraic representation of a linear, time-invariant (LTI) filter in the frequency domain:

The transfer function is also called the

*system function*[60]. Let denote the

*impulse response*of the filter. It turns out (as we will show) that

*the transfer function is equal to the*:

*z*transform of the impulse response*transfer characteristics*of the filter--hence the name. It remains to define ``

*z*transform'', and to prove that the

*z*transform of the impulse response always gives the transfer function, which we will do by proving the

*convolution theorem*for

*z*transforms.

##
The *Z* Transform

The *bilateral*of the discrete-time signal is defined to be

*z*transformwhere is a complex variable. Since signals are typically defined to begin (become nonzero) at time , and since filters are often assumed to be causal,

^{7.1}the lower summation limit given above may be written as 0 rather than to yield the

*unilateral*:

*z*transform(7.2) |

The unilateral

*z*transform is most commonly used. For inverting

*z*transforms, see §6.8. Recall (§4.1) that the mathematical representation of a discrete-time signal maps each integer to a complex number ( ) or real number ( ). The

*z*transform of , on the other hand, , maps every complex number to a new complex number . On a higher level, the

*z*transform, viewed as a

*linear operator*, maps an entire signal to its

*z*transform . We think of this as a ``function to function'' mapping. We may say is the

*z*transform of by writing

*z*transform of a signal can be regarded as a

*polynomial*in , with coefficients given by the signal samples. For example, the signal

*z*transform .

##
Existence of the *Z*
Transform

The *z*transform of a finite-amplitude signal will always

*exist*provided (1) the signal starts at a finite time and (2) it is

*asymptotically exponentially bounded*,

*i.e.*, there exists a finite integer , and finite real numbers and , such that for all . The bounding exponential may even be growing with (). These are not the most general conditions for existence of the

*z*transform, but they suffice for most practical purposes. For a signal growing as , for , one would naturally expect the

*z*transform to be defined only in the region of the complex plane. This is expected because the infinite series

*the region of convergence of the transform of a decaying exponential always includes the unit circle of the plane*. More generally, it turns out that, in all cases of practical interest, the domain of can be

*extended*to include the

*entire complex plane*, except at isolated ``singular'' points

^{7.2}at which approaches infinity (such as at when ). The mathematical technique for doing this is called

*analytic continuation*, and it is described in §D.1 as applied to the

*Laplace transform*(the continuous-time counterpart of the

*z*transform). A point to note, however, is that in the extension region (all points such that in the above example), the signal component corresponding to each singularity inside the extension region is ``flipped'' in the time domain. That is, ``causal'' exponentials become ``anticausal'' exponentials, as discussed in §8.7. The

*z*transform is discussed more fully elsewhere [52,60], and we will derive below only what we will need.

##

Shift and Convolution Theorems

In this section, we prove the highly useful *shift theorem*and

*convolution theorem*for unilateral

*z*transforms. We consider the space of infinitely long, causal, complex sequences , , with for .

### Shift Theorem

The*shift theorem*says that a

*delay*of samples in the time domain corresponds to a

*multiplication by*in the frequency domain:

SHIFT

or, using more common notation,
*z*transform .

*Proof:*

### Convolution Theorem

The*convolution theorem for*states that for any (real or) complex causal signals and ,

*z*transforms*convolution in the time domain is multiplication in the domain*,

*i.e.*,

*Proof:*

##
*Z* Transform of Convolution

From Eq.(5.5), we have that the output from a linear
time-invariant filter with input and impulse response is given
by the *convolution*of and ,

*i.e.*,

where ``'' means convolution as before. Taking the

*z*transform of both sides of Eq.(6.3) and applying the convolution theorem from the preceding section gives

where H(z) is the

*z*transform of the filter impulse response. We may divide Eq.(6.4) by to obtain

*z*transform of an impulse response is equal to the transfer function of the filter, provided the filter is linear and time invariant.

##
*Z* Transform of Difference Equations

Since *z*transforming the convolution representation for digital filters was so fruitful, let's apply it now to the general difference equation, Eq.(5.1). To do this requires two properties of the

*z*transform,

*linearity*(easy to show) and the

*shift theorem*(derived in §6.3 above). Using these two properties, we can write down the

*z*transform of any difference equation by inspection, as we now show. In §6.8.2, we'll show how to

*invert*by inspection as well. Repeating the general difference equation for LTI filters, we have (from Eq.(5.1)) Let's take the

*z*transform of both sides, denoting the transform by . Because is a linear operator, it may be distributed through the terms on the right-hand side as follows:

^{7.3}where we used the superposition and scaling properties of linearity given on page , followed by use of the shift theorem, in that order. The terms in may be grouped together on the left-hand side to get

*z*transform of the difference equation yields

Thus, taking the

*z*transform of the general difference equation led to a new formula for the transfer function in terms of the difference equation coefficients. (Now the minus signs for the feedback coefficients in the difference equation Eq.(5.1) are explained.)

## Factored Form

By the fundamental theorem of algebra, every th order polynomial can be*factored*into a product of first-order polynomials. Therefore, Eq.(6.5) above can be written in

*factored form*as

The numerator roots are called the

*zeros*of the transfer function, and the denominator roots are called the

*poles*of the filter. Poles and zeros are discussed further in Chapter 8.

## Series and Parallel Transfer Functions

The transfer function conveniently captures the*algebraic structure*of a filtering operation with respect to

*series or parallel combination*. Specifically, we have the following cases:

*Transfer functions of filters in series multiply together.**Transfer functions of filters in parallel sum together.*

### Series Case

*series connection*of two filters and . The output from filter 1 is used as the input to filter 2. Therefore, the overall transfer function is

*multiply*together.

### Parallel Case

Figure 6.2 illustrates the*parallel combination*of two filters. The filters and are driven by the

*same*input signal , and their respective outputs and are

*summed*. The transfer function of the parallel combination is therefore

*z*transform to have that .

#### Series Combination is Commutative

Since multiplication of complex numbers is commutative, we have*numerical*performance of the overall filter is usually affected by the ordering of filter stages in a series combination [103]. Chapter 9 further considers numerical performance of filter implementation structures. By the convolution theorem for

*z*transforms, commutativity of a product of transfer functions implies that

*convolution is commutative*:

## Partial Fraction Expansion

An important tool for inverting the*z*transform and converting among digital filter implementation structures is the

*partial fraction expansion*(PFE). The term ``partial fraction expansion'' refers to the expansion of a rational transfer function into a sum of first and/or second-order terms. The case of first-order terms is the simplest and most fundamental:

where

*poles*of the transfer function, and each numerator is called the

*residue*of pole . Equation (6.7) is general only if the poles are

*distinct*. (Repeated poles are addressed in §6.8.5 below.) Both the poles and their residues may be complex. The poles may be found by factoring the polynomial into first-order terms,

^{7.4}

*e.g.*, using the

`roots`function in matlab. The residue corresponding to pole may be found analytically as

when the poles are distinct. The matlab function

`residuez`

^{7.5}will find poles and residues computationally, given the difference-equation (transfer-function) coefficients. Note that in Eq.(6.8), there is always a pole-zero cancellation at . That is, the term is always cancelled by an identical term in the denominator of , which must exist because has a pole at . The residue is simply the

*coefficient*of the one-pole term in the partial fraction expansion of at . The transfer function

*is*, in the limit, as .

### Example

Consider the two-pole filter### Complex Example

To illustrate an example involving complex poles, consider the filter### PFE to Real, Second-Order Sections

When all coefficients of and are real (implying that is the transfer function of a*real*filter), it will always happen that the complex one-pole filters will occur in

*complex conjugate pairs*. Let denote any one-pole section in the PFE of Eq.(6.7). Then if is complex and describes a real filter, we will also find somewhere among the terms in the one-pole expansion. These two terms can be paired to form a

*real second-order section*as follows:

*polar form*as , and the residue as , the last expression above can be rewritten as

*biquads*.

^{7.6}However, the full generality of a biquad section (two poles and two zeros) is not needed because the PFE requires only one zero per second-order term. To see why we must stipulate in Eq.(6.7), consider the sum of two first-order terms by direct calculation:

(7.9) |

Notice that the numerator order, viewed as a polynomial in , is one less than the denominator order. In the same way, it is easily shown by mathematical induction that the sum of one-pole terms can produce a numerator order of at most (while the denominator order is if there are no pole-zero cancellations). Following terminology used for analog filters, we call the case a

*strictly proper transfer function*.

^{7.7}Thus, every strictly proper transfer function (with distinct poles) can be implemented using a parallel bank of two-pole, one-zero filter sections.

### Inverting the Z Transform

The partial fraction expansion (PFE) provides a simple means for inverting the*z*transform of rational transfer functions. The PFE provides a sum of first-order terms of the form

*z*transform of

*z*transform of is simply

*linear combination of sampled complex exponentials*. Recall that a uniformly sampled exponential is the same thing as a

*geometric sequence*. Thus, is a linear combination of geometric sequences. The

*term ratio*of the th geometric sequence is the th pole, , and the

*coefficient*of the th sequence is the th residue, . In the

*improper*case, discussed in the next section, we additionally obtain an

*FIR part*in the

*z*transform to be inverted:

### FIR Part of a PFE

When in Eq.(6.7), we may perform a step of*long division*of to produce an

*FIR part*in parallel with a strictly proper IIR part:

where

`residuez`function (a matlab function for computing a complete partial fraction expansion, as illustrated in §6.8.8 below). An alternate FIR part is obtained by performing long division on the

*reversed*polynomial coefficients to obtain

where is again the order of the FIR part. This type of decomposition is computed (as part of the PFE) by

`residued`, described in §J.6 and illustrated numerically in §6.8.8 below. We may compare these two PFE alternatives as follows: Let denote , , and . (

*I.e.*, we use a subscript to indicate polynomial order, and `' is omitted for notational simplicity.) Then for we have two cases:

*modeling*purposes, since the numerator of the IIR part ( ) can be used to match additional terms in the impulse response after the FIR part has ``died out''. In summary, an arbitrary digital filter transfer function with distinct poles can always be expressed as a parallel combination of

*complex one-pole filters*, together with a parallel FIR part when . When there is an FIR part, the strictly proper IIR part may be delayed such that its impulse response begins where that of the FIR part leaves off. In artificial reverberation applications, the FIR part may correspond to the

*early reflections*, while the IIR part provides the

*late reverb*, which is typically dense, smooth, and exponentially decaying [86]. The

*predelay*(``pre-delay'') control in some commercial reverberators is the amount of pure delay at the beginning of the reverberator's impulse response. Thus, neglecting the early reflections, the order of the FIR part can be viewed as the amount of predelay for the IIR part.

#### Example: The General Biquad PFE

The general second-order case with (the so-called*biquad*section) can be written when as

yielding

giving

### Alternate PFE Methods

Another method for finding the pole residues is to write down the general form of the PFE, obtain a common denominator, expand the numerator terms to obtain a single polynomial, and equate like powers of . This gives a linear system of equations in unknowns , . Yet another method for finding residues is by means of Taylor series expansions of the numerator and denominator about each pole , using l'Hôpital's rule.. Finally, one can alternatively construct a*state space realization*of a strictly proper transfer function (using,

*e.g.*,

`tf2ss`in matlab) and then

*diagonalize*it via a

*similarity transformation*. (See Appendix G for an introduction to state-space models and diagonalizing them via similarity transformations.) The transfer function of the diagonalized state-space model is trivially obtained as a sum of one-pole terms--

*i.e.*, the PFE. In other words, diagonalizing a state-space filter realization implicitly performs a partial fraction expansion of the filter's transfer function. When the poles are distinct, the state-space model can be diagonalized; when there are repeated poles, it can be block-diagonalized instead, as discussed further in §G.10.

### Repeated Poles

When poles are repeated, an interesting new phenomenon emerges. To see what's going on, let's consider two identical poles arranged in parallel and in series. In the parallel case, we have*parallel*are equivalent to a new one-pole filter

^{7.8}(when the poles are identical), while the same two filters in

*series*give a

*two-pole*filter with a repeated pole. To accommodate both possibilities, the general partial fraction expansion must include the terms

#### Dealing with Repeated Poles Analytically

A pole of*multiplicity*has residues associated with it. For example,

and the three residues associated with the pole are 1, 2, and 4. Let denote the th residue associated with the pole , . Successively differentiating times with respect to and setting isolates the residue :

#### Example

For the example of Eq.(6.12), we obtain#### Impulse Response of Repeated Poles

In the time domain, repeated poles give rise to*polynomial amplitude envelopes*on the decaying exponentials corresponding to the (stable) poles. For example, in the case of a single pole repeated twice, we have

*Proof:*First note that

(7.13) |

Note that is a first-order polynomial in . Similarly, a pole repeated three times corresponds to an impulse-response component that is an exponential decay multiplied by a

*quadratic*polynomial in , and so on. As long as , the impulse response will eventually decay to zero, because exponential decay always overtakes polynomial growth in the limit as goes to infinity.

#### So What's Up with Repeated Poles?

In the previous section, we found that repeated poles give rise to polynomial amplitude-envelopes multiplying the exponential decay due to the pole. On the other hand, two*different*poles can only yield a convolution (or sum) of two different exponential decays, with no polynomial envelope allowed. This is true no matter how closely the poles come together; the polynomial envelope can occur only when the poles merge exactly. This might violate one's intuitive expectation of a continuous change when passing from two closely spaced poles to a repeated pole. To study this phenomenon further, consider the convolution of two one-pole impulse-responses and :

The finite limits on the summation result from the fact that both and are causal. Recall the closed-form sum of a truncated geometric series:

(7.15) |

Setting yields

(7.16) |

which is the first-order polynomial amplitude-envelope case for a repeated pole. We can see that the transition from ``two convolved exponentials'' to ``single exponential with a polynomial amplitude envelope'' is perfectly continuous, as we would expect. We also see that the polynomial amplitude-envelopes fundamentally arise from

*iterated convolutions*. This corresponds to the repeated poles being arranged in

*series*, rather than in parallel. The simplest case is when the repeated pole is at , in which case its impulse response is a constant:

^{7.9}

### Alternate Stability Criterion

In §5.6 (page ), a filter was defined to be*stable*if its impulse response decays to 0 in magnitude as time goes to infinity. In §6.8.5, we saw that the impulse response of every finite-order LTI filter can be expressed as a possible FIR part (which is always stable) plus a linear combination of terms of the form , where is some finite-order polynomial in , and is the th pole of the filter. In this form, it is clear that the impulse response always decays to zero when each pole is strictly inside the unit circle of the plane,

*i.e.*, when . Thus, having all poles strictly inside the unit circle is a

*sufficient*criterion for filter stability. If the filter is

*observable*(meaning that there are no pole-zero cancellations in the transfer function from input to output), then this is also a

*necessary*criterion. A transfer function with no pole-zero cancellations is said to be

*irreducible*. For example, is irreducible, while is reducible, since there is the common factor of in the numerator and denominator. Using this terminology, we may state the following stability criterion:

This characterization of stability is pursued further in §8.4, and yet another stability test (most often used in practice) is given in §8.4.1.

### Summary of the Partial Fraction Expansion

In summary, the partial fraction expansion can be used to expand any rational*z*transform

(7.17) |

for , and

for , where the term is optional, but often preferred. For real filters, the complex one-pole terms may be paired up to obtain second-order terms with real coefficients. The PFE procedure occurs in two or three steps:

- When , perform a step of long division to obtain an FIR part and a strictly proper IIR part .
- Find the poles , (roots of ).
- If the poles are distinct, find the residues ,
from
- If there are repeated poles, find the additional residues via
the method of §6.8.5, and the general form of the PFE is

where denotes the number of distinct poles, and denotes the multiplicity of the th pole.

*factoring*the denominator polynomial . This is a dangerous step numerically which may fail when there are many poles, especially when many poles are clustered close together in the plane. The following matlab code illustrates factoring to obtain the three roots, , :

A = [1 0 0 -1]; % Filter denominator polynomial poles = roots(A) % Filter polesSee Chapter 9 for additional discussion regarding digital filters implemented as parallel sections (especially §9.2.2).

### Software for Partial Fraction Expansion

Figure 6.3 illustrates the use of`residuez`(§J.5) for performing a partial fraction expansion on the transfer function

B = [1 0 0 0.125]; A = [1 0 0 0 0 0.9^5]; [r,p,f] = residuez(B,A) % r = % 0.16571 % 0.22774 - 0.02016i % 0.22774 + 0.02016i % 0.18940 + 0.03262i % 0.18940 - 0.03262i % % p = % -0.90000 % -0.27812 - 0.85595i % -0.27812 + 0.85595i % 0.72812 - 0.52901i % 0.72812 + 0.52901i % % f = [](0x0) |

#### Example 2

For the filterwe obtain the output of

`residued`(§J.6) shown in Fig.6.4. In contrast to

`residuez`,

`residued`delays the IIR part until after the FIR part. In contrast to this result,

`residuez`returns

`r=[-24;16]`and

`f=[10;2]`, corresponding to the PFE

(7.22) |

in which the FIR and IIR parts have overlapping impulse responses. See Sections J.5 and J.6 starting on page for listings of

`residuez`,

`residued`and related discussion.

B=[2 6 6 2]; A=[1 -2 1]; [r,p,f,m] = residued(B,A) % r = % 8 % 16 % % p = % 1 % 1 % % f = % 2 10 % % m = % 1 % 2 |

#### Polynomial Multiplication in Matlab

The matlab function`conv`(

*convolution*) can be used to perform

*polynomial multiplication*. For example:

B1 = [1 1]; % 1st row of Pascal's triangle B2 = [1 2 1]; % 2nd row of Pascal's triangle B3 = conv(B1,B2) % 3rd row % B3 = 1 3 3 1 B4 = conv(B1,B3) % 4th row % B4 = 1 4 6 4 1 % ...The matlab

`conv(B1,B2)`is identical to

`filter(B1,1,B2)`, except that

`conv`returns the

*complete*convolution of its two input vectors, while

`filter`truncates the result to the length of the ``input signal''

`B2`.

^{7.10}Thus, if

`B2`is zero-padded with

`length(B1)-1`zeros, it will return the complete convolution:

B1 = [1 2 3]; B2 = [4 5 6 7]; conv(B1,B2) % ans = 4 13 28 34 32 21 filter(B1,1,B2) % ans = 4 13 28 34 filter(B1,1,[B2,zeros(1,length(B1)-1)]) % ans = 4 13 28 34 32 21

#### Polynomial Division in Matlab

The matlab function`deconv`(

*deconvolution*) can be used to perform

*polynomial long division*in order to split an improper transfer function into its FIR and strictly proper parts:

B = [ 2 6 6 2]; % 2*(1+1/z)^3 A = [ 1 -2 1]; % (1-1/z)^2 [firpart,remainder] = deconv(B,A) % firpart = % 2 10 % remainder = % 0 0 24 -8Thus, this example finds that is as written in Eq.(6.21). This result can be checked by obtaining a common denominator in order to recalculate the direct-form numerator:

Bh = remainder + conv(firpart,A) % = 2 6 6 2The operation

`deconv(B,A)`can be implemented using

`filter`in a manner analogous to the polynomial multiplication case (see §6.8.8 above):

firpart = filter(B,A,[1,zeros(1,length(B)-length(A))]) % = 2 10 remainder = B - conv(firpart,A) % = 0 0 24 -8That this must work can be seen by looking at Eq.(6.21) and noting that the impulse-response of the remainder (the strictly proper part) does not begin until time , so that the first two samples of the impulse-response come only from the FIR part. In summary, we may conveniently use convolution and deconvolution to perform polynomial multiplication and division, respectively, such as when converting transfer functions to various alternate forms. When carrying out a partial fraction expansion on a transfer function having a numerator order which equals or exceeds the denominator order, a necessary preliminary step is to perform long division to obtain an FIR filter in parallel with a strictly proper transfer function. This section describes how an FIR part of any length can be extracted from an IIR filter, and this can be used for PFEs as well as for more advanced applications [].

## Problems

See`http://ccrma.stanford.edu/~jos/filtersp/Transfer_Function_Analysis_Problems.html`.

**Next Section:**

Frequency Response Analysis

**Previous Section:**

Time Domain Digital Filter Representations