Complex Low-Pass Filters

PETER KISS, VLADIMIR PRODANOV∗ AND JACK GLAS

Department of Communications Circuits Research, Agere Systems, 4 Connell Drive, Berkeley Heights, NJ 07922, USA
E-mail: [email protected]; [email protected]; [email protected]

Abstract. Zero-if transceivers suffer from the imbalance of the I and Q paths. By using a complex low-pass filter
topology instead of a conventional pair of real low-pass filters, this imperfection can be reduced. Both analytical
and numerical analysis show that the proposed technique is significantly more robust to circuit imperfections than
the traditional architecture.

Key Words: complex filters, poliphase filters, sensitivity analysis, I/Q imbalance, direct conversion, zero if

1. Motivation low-pass filters outperform the pair of real low-pass fil­

ters from I/Q imbalance point of view—without adding
The performance of practical direct-conversion (zero- significantly to the hardware complexity.
if) transceivers (Fig. 1(a)), among other imperfections,
suffers from I/Q imbalance caused by the mismatch of
the mixers and the imperfect quadrature signals from 2. Real Filters
the local oscillators [1,2]. In addition, the mismatch be­
tween the frequency responses of the two real low-pass In a conventional zero-if architecture [1,2], LPF1 and
filters (LPF1 and LPF2 in Fig. 1(a)) also contributes to LPF2 form a two-input two-output linear network with
.
the I/Q imbalance of the receiver, causing performance complex input xc (t) = x1 (t) + j x2 (t) and complex out­
.
degradation. When zero-if topology is proposed for put yc (t) = y1 (t) + j y2 (t) (Fig. 1(a), for details about
wide-band applications with more stringent require- complex signals refer to Appendix A). If the transfer
ments (e.g., 802.11a where SNR = ∼ 30 dB) then this functions of LPF1 and LPF2 are defined as H1 (ω) =
B1 (ω)
distortion needs to be taken into account.
A1 (ω)
and H2 (ω) = AB22 (ω)
(ω)
, then
In this paper a single complex low-pass filter (LPFc
in Fig. 1(b)) is proposed to reduce the I/Q imbalance A2 (ω)B1 (ω) + A1 (ω)B2 (ω)
due to the two-path filtering. Complex filters has been Yc (ω) = (X 1 (ω)
2A1 (ω)A2 (ω)
discussed for many years [3]. In the late 1960s, ana­ A2 (ω)B1 (ω) − A1 (ω)B2 (ω)
log polyphase filters were proposed for single-sideband + j X 2 (ω)) +
2A1 (ω)A2 (ω)
generators [4] and receivers [5]. About a decade later,
their synthesis and analysis led to new developments × (X 1 (ω) − j X 2 (ω))
[6–8]. Currently they are often used in low-if receiver = Hcm (ω)X c (ω) + Hd f (ω)X c∗ (−ω) (1)
ICs, e.g., [9–11].
Complex low-pass filters are a particular case of the Equation (1) shows that the input complex signal xc (t)
popular complex band-pass filters [3–11]; however, as is processed in a parallel fashion by h cm (t) and h d f (t)
far as the authors are aware, the technical literature does (Fig. 2(a)). The common component of H1 (ω) and
not talk about them. Here, after a brief review of basic H2 (ω) forms Hcm (ω) which gives the desired (direct)
concepts, a detailed sensitivity analysis of such filters output Hcm (ω) · X c (ω). However, if H1 (ω) and H2 (ω)
will be provided. This will demonstrate that complex are not identical, then a nonzero Hd f (ω) contributes to
a leakage (undesired or difference) output component
∗ Corresponding author. Hd f (ω) · X c∗ (−ω).
 Fig. 1. Quadrature direct conversion receiver with (a) two real low-pass filters; (b) one complex low-pass filter.

Fig. 2. (a) Time-domain and (b) frequency-domain model of an imperfect two-path low-pass filter. Imperfect filtering of a complex (c) positive-
frequency and (d) negative-frequency input tone.

For example, if a complex positive-frequency tone the desired signal and the undesired image, respec­
at ω0 undergoes an imperfect two-path filtering opera­ tively. The imperfect two-path filtering, expressed by
tion, then the complex output will contain, besides the equation (1), means that a fraction of the positive-
desired component at ω0 , a leakage component at −ω0 frequency signal X p (ω) will be transformed into a
(Fig. 2(c)). Similarly, a complex input tone at −ω0 will negative-frequency signal X ∗p (−ω) which leaks on top
leak into ω0 (Fig. 2(d)). of X n (ω) and distorts it (Fig. 3). Similarly, a fraction
Let X p (ω) and X n (ω) denote the positive and of X n∗ (−ω) distorts X p (ω).
negative frequency content of X c (ω), respectively Note that this distortion occurs even if the complex
(Fig. 3(b)). Usually, X p (ω) and X n (ω) correspond to local oscillator signal LOc is a single complex tone
Fig. 3. Signal processing: (a) rf input signal; (b) ideal zero-if I/Q mixing; (c) filtering: desired and leakage signals; (d) output signal distorted
by imperfect I/Q filtering.

at −ωlo , as it was assumed in Fig. 3(b). In practical without adding significantly to the hardware complex­
situations, when LOc is not a perfect quadrature, then ity of the filter.
the effects of both imperfections add. A complex low-pass filter is a particular case of
the popular complex band-pass filter when ωi f = 0
(Fig. 4(b)). Note that every complex pole is doubled and
one of them is cancelled by a complex zero. The ideal
3. Complex Filters response of such a filter can be designed to be identical
to the ideal response of the pair of LPF1 and LPF2 .
A complex filter is a two-input two-output linear net­
work which frequency response is not necessarily sym­
metrical with respect to dc (ω = 0). Its gain and phase 3.1. “Single” Complex Pole
responses are functions both of the frequency and the
relative phase difference of the two real inputs x1 and The circuit implementation of complex filters in­
x2 (Fig. 1(b)). volves realizing non-complex-conjugate (single) com­
As an example, a fourth-order all-pole band-pass plex poles. This can be achieved by a pair of complex
complex filter with bandwidth BW centered around conjugate poles out of which one is cancelled by a sin­
ωi f is shown in Fig. 4(a). Since Hc (ω)|ω∈BW = ∼ 0 dB gle complex zero [6] (Fig. 4).
and Hc (−ω)|ω∈BW � 0 dB, the complex band-pass fil­ A “single” complex pole can effectively be imple­
ter provides image rejection in addition to filtering— mented using two integrators in a feedback loop with
 Fig. 4. Frequency response and pole-zero constellation for a 4th-order (a) complex band-pass filter; (b) complex low-pass filter.

Fig. 5. First-order complex filter implementing a “single” complex pole.

two inputs and two outputs [8] (Fig. 5(a)). The complex Due to circuit imperfections, usually a11 =� a22 =
� a
output becomes and a12 =� a21 =� b, so the pole-zero cancellation does
not hold. Equation (2) can be written as follows (similar
. s + a22 + ja21 s + a11 + ja12
Yc (s) = X 1 (s) + j X 2 (s) to [12, p. 58])
D(s) D(s)
(2) s + a11 +a 22
+ j a12 +a 21

Yc (s) = 2 2
(X 1 (s) + j X 2 (s))
D(s)
where D(s) = s 2 + (a11 + a22 )s + a11 a22 + a12 a21 (for a11 −a22
+ j a12 −a 21

details of obtaining equation (2) refer to Appendix B). + 2 2

(X 1 (s) − j X 2 (s))
In ideal case, i.e., a11 = a22 = a and a12 = a21 = b, D(s)
equation (2) becomes = Hcm (s)X c (s) + Hd f (s)X c∗ (s ∗ ) (4)
⇒ Yc (ω) = Hcm (ω)X c (ω) + Hd f (ω)X c∗ (−ω) (5)
s + a + jb
Ycid (s) = (X 1 (s) + j X 2 (s)) Since this last result is identical with equation (1), a
(s + a + jb)(s + a − jb)
1 complex filter behaves in the same way as a pair of
= X c (s) = Hid1 (s)X c (s) (3) real filters, so it can be modeled by Fig. 2(a). Also,
s + a − jb
a mismatched complex filter causes distortion as ex­
Equation (3) shows that the first-order ideal complex plained earlier and illustrated by Fig. 3. However, note
filter Hid1 (s) implements a “single” complex pole p = that Hcm (ω) and Hd f (ω) have real coefficients in equa­
−a + jb, based on a perfect pole cancellation by the tion (1), but they have complex ones in equation (5).
zero z = −a − jb. Therefore, Hcm (ω) and Hd f (ω) are even functions in
Fig. 6. First-order complex band-pass filter with f i f = 13 MHz and σn = 5% mismatch; z 1 = −a11 − ja12 , p1 = −a11 + ja12 , z 2 = −a22 −
ja21 , and p2 = −a22 + ja21 .

ω in equation (1), but they are asymmetrical in ω in quadrature complex signal, i.e., xc (t) = A cos(ω0 t) +
equation (5). j A sin(ω0 t), was applied to the input of the filter. The
The coefficients a11 , a12 , a21 and a22 in Fig. 5(a) are spectrum of the resulting complex output yc (t) was
realized by various circuit elements depending on their measured at ω0 and −ω0 , providing the values for
implementations (e.g., passive R-C [5], active R-C Hcm (ω0 ) and Hd f (−ω0 ), respectively. The experiment
[8], gm -C [11], etc.). Here, a normally-distributed error was performed for all the range of frequencies of in­
with 1% variance was considered for each coefficient; terest. In addition, the transfer functions Hcm (ω) and
the errors were assumed to be uncorrelated. Therefore, Hd f (ω) were calculated based on equation (4). The
the magnitude and distribution of the errors need to be numerical and analytical results were identical which
tailored to the specifics of the implementation. proved the validity of the model described by equa­
The simulated behavior of an imperfect versus per­ tion (4). Equations (6) and (7), defined and discussed
fect “single” complex pole is shown in Fig. 6 (similar later, were verified and validated by a similar simula­
plots can be found in [12, p. 59]). The ideal pole-zero tion procedure.
constellation presents a perfect pole-zero cancellation,
and the frequency response Hid1 (ω) looks as expected. 4. Cascade of Filters
In the presence of 5% errors1 the poles p1 and p2 of the
filter move away from the ideal value of p which leads High-order transfer functions can be realized by a
to a nonzero Hd f (s). However, the pole-zero cancella­ cascade arrangement of elementary building blocks.
tion within Hcm (s) occurs at a high degree (for details For example, a fourth-order all-pole complex low-
refer to Appendix C). pass filter (CLPF) can be built from four “single”
Note that the simulations were performed us­ complex poles, while a similar pair of real low-
ing a black-box approach. In this method a perfect pass filters (RLPF) uses four biquads (Fig. 7). Both
Fig. 7. Two possible implementation of a fourth-order complex low-pass filter.
 Fig. 8. Parallel model for a cascade of imperfect complex poles.

implementations need eight integrators, thus the hard­ 4.1. Comparative Sensitivity Analysis
ware complexity is roughly the same. Note that for a
pair of RLPFs there is no interaction between the indi­ Replacing a pair of RLPFs LPF1 and LPF2 with a CLPF
vidual I and Q stages but at the global output (Fig. 7). in a direct-conversion receiver (Figs. 1(a) vs. (b)) is
Therefore, its leakage, i.e., Hd f (ω), depends only on motivated by the expected increased robustness of the
the global transfer function of LPF1 and LPF2 , and latter. A comparative sensitivity analysis will be pre­
equation (1) holds for any order of the filters. sented in the following.
On the other hand, for a cascade of CLPFs As an example, two imperfect RLPF and CLPF are
the desired (direct) and undesired (leakage) signal compared in Fig. 9. Both are fourth-order 8.5-MHz
components interact at the output of every stage Chebyshev all-pole filters with a pass-band ripple of
(Fig. 8). Therefore, each of the four stages processes R p = 1 dB. They are affected by a normally distributed
its complex input and provides desired and undesired error with σn = 5%, so their poles lay in clusters around
output, according to equation (5) and illustrated by sim­ the ideal locations. For the filters (i.e., LPF1 , LPF2 and
ulation results in Fig. 6. Due to this leakage mechanism, CLPF) cascade (as opposed to, e.g., ladder) implemen­
the image signal component of the input leaks into the tations were assumed. Note that Hcm (ω) and Hd f (ω)
desired signal, and vice versa. Moreover, the signal are even functions in ω for RLPFs, but they are asym­
may leak more than once contributing accordingly to metrical in ω for CLPFs.
the global transfer functions. One can define the average image-rejection ratio
Therefore, over a bandwidth BW as
� � �
� �2
Hcm (s) = Hcm4 (s)Hcm3 (s)Hcm2 (s)Hcm1 (s) IMR = 10 log10 � Hcm (ω) � dω (dB) (8)
� �
ω∈BW Hd f (ω)
+ Hd f 4 (s)Hd∗f 3 (s ∗ )Hcm2 (s)Hcm1 (s) + · · ·
(6) which shows how effectively a complex filter passes
signal inputs while rejecting image inputs [12, p. 59].
∗ The RLPF and CLPF lead to IMRr = 15.8 dB and
Hd f (s) = Hd f 4 (s)Hcm3 (s ∗ )Hcm2
∗
(s ∗ )Hcm1
∗
(s ∗ )
IMRc = 23.0 dB, respectively. Therefore, the complex
∗
+ Hcm4 (s)Hd f 3 (s)Hcm2 (s ∗ )Hcm1
∗
(s ∗ ) filter is 7.2 dB better than the pair of real filters in
∗
+ Hcm4 (s)Hcm3 (s)Hd f 2 (s)Hcm1 (s ∗ ) this example. However, these are just partial results.
+ Hcm4 (s)Hcm3 (s)Hcm2 (s)Hdf 1 (s) + · · · In order to draw general conclusions, the experiment
presented in Fig. 9 was repeated for several mismatch
(7)
states (as in a Monte-Carlo type analysis) and the results
were processed statistically. Moreover, the effect of the
Hcm (ω) and Hd f (ω) contain even and odd numbers sequence of stages was investigated—presented next.
of time-domain2 complex conjugate operations on the
input signal xc (t), respectively. According to equa­
tion (7), the leakage of CLPFs is given by a combi­ 4.2. Sequence of Complex Poles in CLPFs
nation of the individual transfer functions. Therefore,
there is a degree of freedom to sequence the individual The sequence of the stages plays a significant role in
stages in order to minimize the global Hd f (ω). the complex filter’s performance. From equation (4)
 Fig. 9. RLPF versus CLPF for N = 4th order.

Fig. 10. Simulation example for a cascade of four complex stages.

results that the magnitude of one stage’s leakage

√ is pro- most sensitive pole(s) needs to be minimized. This is
portional to the quality factor Q, i.e., Q = 2a
1
a 2 + b2 , illustrated for a fourth-order filter in Fig. 10. The leak­
of the pole it implements. Should the stages, therefore, age of the pole 2 is given by Hd f 2 (ω) which sees a
∗
be ordered in a reverse sequence of their pole Q-s? leakage gain of Hcm3 (−ω)Hcm1 (ω)Hcm4 (ω). This gain
In order to minimize the total leakage at the output is evenly distributed for positive and negative frequen­
of a cascaded CLPF, the “leakage gain” seen by the cies, so it has the lowest possible average value. The
 Fig. 11. The effect of ordering the poles in CLPF for N = 4.

same is true for the other high-Q pole 1. Next, this Table 1. All 24 permutations of four poles.
intuitive reason will be verified by numerical methods.
n code(n) n code(n) n code(n) n code(n)
Figure 11 presents the results of a statistical analysis.
The variable n indicates the sequence of stages (each 1 1-2-3-4 7 2-1-3-4 13 3-1-2-4 19 4-1-2-3
stage implements an imperfect complex pole); the 2 1-2-4-3 8 2-1-4-3 14 3-1-4-2 20 4-1-3-2
3 1-3-2-4 9 2-3-1-4 15 3-2-1-4 21 4-2-1-3
behavior of all 4! = 24 possible permutations are
4 1-3-4-2 10 2-3-4-1 16 3-2-4-1 22 4-2-3-1
shown. The poles are labeled 1 . . . N , sequenced in 5 1-4-2-3 11 2-4-1-3 17 3-4-1-2 23 4-3-1-2
a decreasing order of their Q-s, but first the pole on 6 1-4-3-2 12 2-4-3-1 18 3-4-2-1 24 4-3-2-1
the positive side of ω comes, then that one on the
negative side of ω follows (Fig. 4(b)). The pole se­
quences corresponding to different n-s are given in for n = 19 (CLPF) and n = 25 (RLPF), respectively.
Table 1. For example, code(15) = 3-2-1-4 means that Note that CLPF has a larger mean and lower variance
pole 3 is implemented in the first stage followed by than RLPF.
pole 2, pole 1, and pole 4, like in Fig. 10. In Fig. 8 the Based on the IMR performance of CLPFs shown in
poles were arranged according to n = 1, i.e., code(1) = Fig. 11, three categories of CLPFs can be clearly iden­
1-2-3-4. n = 25 is for RLPF. For each sequence n, a tified: “best” (n ∈ {15, 19}), “mediocre” (n ∈ {1, 2,
set of 10000 normally distributed (σn = 1%) random 5–10, 16–18, 20, 23, 24}), and “worst” (n ∈ {3, 4,
mismatch states were simulated which error-bar (mean 11–14, 21, 22}). In the “best” group the poles follow a
value and variance) is shown in Fig. 11. On the lower shoestring pattern. There are only two such sequences
part of this figure, two histograms of IMR are shown possible, i.e., code(15) = 3-2-1-4 and code(19) =
 Fig. 12. Shoestring patterns of poles for (a) N = 4; (b) N = 6.

4-1-2-3 (Fig. 12(a)). Indeed, these sequences minimize The histograms of Fig. 11 allow determining the
the “leakage gain” seen by the most sensitive poles. Fi­ yield of such filters (Fig. 13). These curves reveal a
nally, Fig. 11 shows that the best sequenced CLPFs more dramatic comparison. For example, if an appli­
achieve about 3 dB larger IMR compared to RLPFs cation requires an IMR of 30 dB, then using RLPFs vs.
for N = 4. Note that this result depends on the highest CLPFs will result of about 20% lower yield. Moreover,
value of the pole Q-s; for the filter in Fig. 9, Q max = 3.5. if a mass production needs to achieve a yield not lower
If a ripple of 3 dB is assumed, which boosts up the Q max than 90%, using CLPFs vs. RLPFs provides an excess
to 5.5, then �IMR becomes 4.1 dB. Therefore, the pro­ of 4 dB of IMR.
posed technique is more effective for high-Q (i.e., more The benefits of using CLPF over RLPF improve
selective) filters. when the filter’s order increases; note that high-order

Fig. 13. Comparative yield curves for N = 4 (data from Fig. 12).
 Fig. 14. CLPF versus RLPF in function of N .

Table 2. CLPF versus RLPF in function of N . In summary, the intuitive and statistical analysis pre­
sented in this section demonstrated that it is possible to
N , order 2 3 4 5 6 7 8
predict a priori the most robust CLPF topology. Also, it
Q max 0.9 2.0 3.5 5.5 8.0 10.9 14.2 turns out that the CLPFs are significantly less sensitive
IMRc (dB) 41.0 37.8 37.5 35.3 35.4 33.5 33.6 to circuit imperfections than RLPFs.
IMRr (dB) 41.5 37.6 34.5 31.8 29.3 27.9 26.6
�IMR (dB) −0.5 0.2 3.0 3.5 6.1 5.6 7.0
5. Conclusions

filters are more likely to use high-Q poles. This is sum­ The proposed complex low-pass filter is a novel topol­
marized in Table 2 and Fig. 14, e.g., the improvement ogy suitable for direct-conversion transceivers. De­
is about 7 dB for N = 8. It is interesting to note that the tailed analytical and numerical analysis were pre­
number of poles, thus the selectivity, of an odd-order sented. In order to reach maximal robustness for a com­
CLPF can be incremented for “free,” while this costs plex low-pass filter, its stages should be ordered in an
IMR degradation in the case of RLPFs (Fig. 14). a priori predictable shoestring pattern. It turns out that
It was verified by simulations that in the case of high- the complex low-pass filters are several dB-s more ro­
order filters, the above-described shoestring criteria for bust to circuit imperfections than the traditionally used
choosing the best sequence for CLPF remains valid. pair of real low-pass filters. Moreover, the proposed
For example, for N = 6 the two best sequences should technique is even more effective for high-order high-Q
be 5-4-1-2-3-6 and 6-3-2-1-4-5 (Fig. 12(b)). For odd- filters.
order filters, the position of the real pole does not mat­
ter much since its leakage is small and its response is
symmetrical in respect to dc. However, it is preferred to Appendix
place the real pole in the middle or at the extremes (i.e.,
beginning or end) of the cascade in order to keep the A. Complex Exponential
shoestring pattern symmetrical. For N = 2 and N = 3
the CLPF is unbalanced, and it performs similar to the The complex exponential Ae jω0 t can be considered
RLPF (Table 2). as the mathematical model of a perfect quadrature
Fig. 15. Complex exponentials formed by a pair of (a) forward-quadrature signals, X c (ω) = F {A cos(ω0 t) + j A sin(ω0 t)}; (b) reverse-
quadrature signals, X c (ω) = F{A cos(ω0 t) − j A sin(ω0 t)}.

oscillation. It is generated by two real signals x1 and ↓→ . x c can be represented by a reverse-quadrature or

x2 (Fig. 15(a)) which oscillate with the same an­ reverse-rotating phasor, which rotates clockwise with
gular frequency ω0 and amplitude A. They should −ω0 (or counter-clockwise with ω0 ), and forms the
be in a quadrature relationship,3 that is, x2 (φ2 (t)) = complex exponential Ae− jω0 t (Fig. 15(b))
x1 (φ1 (t) − π2 ), for ∀t > 0; x 2 is the Hilbert transform 
x1 (t) = A cos(ω0 t) ↔ X 1 (ω) = Aπ {δ(ω

[14, Section 11.4] of x1 , and (x1 , x2 ) can be called 

 + ω0 ) + δ(ω − ω0 )}
a forward-quadrature pair, ↑→ . From the two real sig­ x2 (t) = −A sin(ω0 t) ↔ X 2 (ω) = (Aπ/j)
nals x1 and x2 , the complex signal xc is obtained by 

. 
 ×{δ(ω + ω0 ) − δ(ω − ω0 )}
the xc (t) = x1 (t) + j x2 (t) operation. In other words,  .
xc (t) = x1 (t) + j x2 (t) ↔ X c (ω) = 2Aπ δ(ω + ω0 )
Re{xc (t)} = x1 (t) and Im{xc (t)} = x2 (t). xc is also
(10)
called the analytic part [15, p. 119] of x1 . It is con­
venient to represent xc as a forward-quadrature or Note that the Fourier transform X c (ω) is a single Dirac
forward-rotating phasor, which rotates clockwise with delta function at −ω0 (Fig. 15(b)).
ω0 , and x1 always leads x2 with π2 radians satisfying the
quadrature relationship. The time-frequency represen­
tation of these signals is given graphically in Fig. 15(a), B. “Single” Complex Pole
and analytically below
 An effective implementation of a “single” complex
x1 (t) = A cos(ω0 t) ↔ X 1 (ω) = Aπ{δ(ω
 pole can be done using two integrators within a feed­


 + ω0 ) + δ(ω − ω0 )} back loop with two inputs and two outputs (Fig. 5(a)).
x2 (t) = A sin(ω0 t) ↔ X 2 (ω) = −(Aπ/j) This diagram is redrawn in Fig. 5(b) which empha­



 ×{δ(ω + ω0 ) − δ(ω − ω0 )} sizes on the negative feedback loop used. The amount
 .
xc (t) = x1 (t) + j x2 (t) ↔ X c (ω) = 2Aπδ(ω − ω0 ) of feedback [16, Section 8.1] is given by
(9) a12 a21
“1 + β A” = 1 +
(s + a11 )(s + a22 )
Note that xc is a double-wired signal and its Fourier
transform X c (ω) is a single Dirac delta function at ω0 s 2 + (a11 + a22 )s + a11 a22 + a12 a21
=
(Fig. 15(a)). (s + a11 )(s + a22 )
When x1 lags x2 with π2 radians (or x2 leads x1 with D(s)
π = (11)
2
), (x1 , x2 ) can be called a reverse-quadrature pair, (s + a11 )(s + a22 )
The voltage-gain parameters4 H of this circuit are the The zero z cm of Hcm (s) is given by
following Hcm (s) = 0 ⇒ s = z cm
 �
 . Y1 (s) �� 1 1 s + a22 a11 + a22 a12 + a21

 H11 = = · = =− −j (16)

 �
X 1 (s) x2 =0 s + a11 1 + β A D(s) 2 2



 �

 . Y2 (s) �� a21 1 In order to compare the location of p2 and z cm , let us


 H12 =
 � = · assume that a21 = a12 (1 + �b ) and a22 = a11 (1 + �a );

 X 1 (s) x2 =0 (s + a11 )(s + a22 ) 1 + β A

 equations (15) and (16) become

 a21

 =  �

 a11 + a22
D(s) 
 1
� 
 p2 = − − j a12 a21 − (a11 − a22 )2

 2 4


 . Y1 (s) �� −a12 1 
 �

 H21 = � = · 
 a11 + a22 a11 2

 X 2 (s) x1 =0 (s + a11 )(s + a22 ) 1 + β A  ∼
=− − ja12 1 + �b + �

 2 4a12 a

 −a12
 



 = z = − a11 + a22 − j a12 + a21 = − a11 + a22


 D(s) 
 cm

 
 2� � 2 2

 � 
 �b


 . Y2 (s) �� 1 1 s + a11  − ja12 1 +
 H22 = � = · = 2
X (s)
2 x1 =0s +a 1+βA
22 D(s)
(17)
(12)
Since for small values of �, �2 is negligible, and 1 + �2
which can be written in the following format √
� � � � is the Maclaurin’s series expansion of 1 + �, equa­
. H11 H12 1 s + a22 a21 tion (17) shows that z cm ∼ = p2 , that is, the pole-zero
H (s) = = ·
H21 H22 D(s) −a12 s + a1 cancellation in Hcm (s) still occurs at a very good ap­
(13) proximation. In conclusion, the response of Hcm (ω) is
expected to be a “shifted” version of Hid (ω). On the
Therefore, the complex output becomes other hand, Hd f (s) holds the same poles as Hcm (s), but
.
Yc (s) = L{y1 (t) + j y2 (t)} = Y1 (s) + jY2 (s) it has no finite zeros and it has a small gain—directly
proportional with the magnitude of �a and �b .
= (H11 (s)X 1 (s) + H21 (s)X 2 (s))
+ j (H12 (s)X 1 (s) + H22 )(s)X 2 (s)
Acknowledgments
= (H11 (s) + j H12 (s))X 1 (s) + (H22 (s)
− j H21 (s)) j X 2 (s) Useful discussions with Dr. Mihai Banu of Agere Sys­
s + a22 + ja21 tems are gratefully acknowledged.
= X 1 (s)
D(s)
s + a11 + ja12 Notes
+ j X 2 (s) (14)
D(s)
1. In Figs. 6 and 9 the errors were kept larger than the practical value
of 1% in order to show meaningful pole-zero constellations.
2. Note that the Fourier-transform pair of x ∗ (t) is X ∗ (−ω), since
C. Imperfect Pole-Zero Cancellation
L{x ∗ (t)} = X ∗ (s ∗ ) [13, p. 691].
3. Trigonometrical reminder: sin(φ + π2 ) = cos(φ), cos(φ − π2 ) =
Let us take a closer look to Hcm (s) and Hd f (s) given sin(φ), sin(φ − π2 ) = −cos(φ), cos(φ + π2 ) = −sin(φ), sin(φ ±
by equation (4). Their poles are given by the roots of π ) = −sin(φ), and cos(φ ± π ) = −cos(φ).
D(s). Since all the coefficients of D(s) are positive real 4. The H parameters defined by equation (12) are different from the
traditional h parameters [16, Appendix B] of two-port circuits.
numbers, it has complex conjugate roots, i.e., D(s) =
(s − p)(s − p ∗ ) = (s − p1 )(s − p2 ), where
a11 + a22 References
D(s) = 0 ⇒ s = p1,2 = −
2
1�
1. Abidi, A. A., “Direct-conversion radio transceivers for digital
± (2 j)2 a12 a21 + (a11 − a22 )2 (15) communications.” IEEE Journal of Solid-State Circuits 30(12),
2 pp. 1399–1410, 1995.
 2. Razavi, B., “Design considerations for direct-conversion re­ Technical University of Timişoara, Romania, in 1994,
ceivers.” IEEE Journal of Solid-State Circuits 44(6), pp. 428– 1995 and 2000, respectively. From 1998 to 2000 he was
435, 1997.
a research scholar at Oregon State University, Corval­
3. Gabor, D., “Theory of communication.” Journal of the Institu­
tion of Electrical Engineers 93(Part III), pp. 429–457, 1946. lis, OR, working on correction techniques for fast and
4. Gingell, M. J., “A symmetrical polyphase network.” U.S. Patent accurate and large-bandwidth delta-sigma converters.
3,559,042, 1968. Since 2001 he has been with Agere Systems (formerly
5. Gingell, M. J., “The synthesis and application of polyphase part of Bell Laboratories) in Murray Hill, NJ, deal­
networks with sequence asymmetric properties.” Ph.D.
ing with data converters and analog filters for wireless
Thesis, Faculty of Engineering, University of London.
http://users.vnet.net/gingell/, 1975. application. His past work involved adaptive fuzzy sys­
6. Lang, G. R. and Brackett, P. O., “Complex analogue filters,” in tems and image processing.
Proceedings of the European Conference on Circuit Theory and
Design, pp. 412–419, 1981.
7. Snelgrove, W. M. and Sedra, A. S., “State-space synthesis of
complex analog filters,” in Proceedings of the European Con­
ference on Circuit Theory and Design, pp. 420–424, 1981.
8. Sedra, A. S., Snelgrove, W. M. and Allen, R., “Complex analog
bandpass filters designed by linearly shifting real low-pass pro­
totypes,” in Proceedings of the IEEE International Symposium
on Circuits and Systems, pp. III.1223–III.1226, 1985.
9. Crols, J. and Steyaert, M. S., “A single-chip 900 MHz CMOS
receiver front-end with a high performance low-IF topology.”
IEEE Journal of Solid-State Circuits 30(12), pp. 1483–1492,
1995.
10. Jantzi, S. A., Martin, K. W. and Sedra, A. S., “Quadrature band-
pass delta-sigma modulation for digital radio.” IEEE Journal of
Solid-State Circuits 32(12), pp. 1935–1950, 1997.
11. Prodanov, V., Palaskas, G., Glas, J. and Boccuzzi, V., “A CMOS
AGC-less IF strip for Bluetooth,” in IEEE European Solid-State Vladimir Prodanov is an analog circuit designer
Circuits Conference, Digest of Technical Papers, pp. 488–491, with Communication Circuits Dept., Agere Sys­
2001. tems (formerly Silicon Circuits Research Dept., Bell
12. Jantzi, S. A., “Quadrature bandpass delta-sigma modulation for Laboratories, Lucent Technologies). Mr. Prodanov is
digital radio.” Ph.D. Thesis, Electrical and Computer Engineer­
ing, University of Toronto, 1997.
also an adjunct professor at Columbia University, New
13. Oppenheim, A., Willsky, A. S. and Nawab, S. H., Signals and York City, where he teaches analog and RF circuit
Systems. Prentice-Hall, New Jersey, 1996. design. Mr. Prodanov received B.S. degree from Tech­
14. Oppenheim, A. and Schafer, R. W., Discrete-Time Signal Pro­ nical University, Sofia, Bulgaria, and M.S. and Ph.D.
cessing. Prentice-Hall, New Jersey, 1998. from State University of New York at Stony Brook, in
15. Papoulis, A., Signal Analysis. McGraw-Hill, New York, 1977.
16. Sedra, A. S. and Smith, K. C., Microelectronic Circuits. Oxford
1991, 1995 and 1997, respectively.
University Press: New York, 1998.

Peter
´ Kiss received the Engineer’s, M.S. and Jack Glas received M.Sc. (with honors) and
Ph.D. degrees, all in electrical engineering, from the Ph.D. degrees in electrical engineering from Delft
University of Technology in the Netherlands in 1992 Some of the systems he has been working on include
and 1996, respectively. After graduation he joined Bell Cellular systems (such as GSM) as well as Wire­
Laboratories (Lucent Technologies) where he contin­ less LAN systems (such as 802.11a/b/g, Bluetooth).
ued research in the areas of RF System Design and Currently he is part of the Wireless Circuits Research
Baseband Signal Processing algorithms. His focus has Department of Agere Systems where he is looking into
been on advanced architectures that reduce cost and next generation Wireless LAN systems.
power consumption while maintaining performance.