324 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 33, NO.

3, MARCH 1998

Cyclostationary Noise Analysis of Large

RF Circuits with Multitone Excitations
Jaijeet Roychowdhury, David Long, and Peter Feldmann, Senior Member, IEEE

Abstract—This paper introduces a new, efficient technique for this paper can analyze noise efficiently in large circuits with
analyzing noise in large RF circuits subjected to true multitone multitone large-signal inputs.
excitations. Noise statistics in such circuits are time-varying, Being periodic or quasi-periodic, the statistics of cyclo-
hence cyclostationary stochastic processes, characterized by har-
monic power spectral densities (HPSD’s), are used to describe stationary processes can be expressed in Fourier series. The
noise. HPSD’s are used to devise a harmonic-balance-based present technique is formulated in terms of the coefficients
noise algorithm with the property that required computational of the Fourier series, termed cyclostationary components.
resources grow almost linearly with circuit size and nonlinearity. The cyclostationary components, which the new algorithm
Device noises with arbitrary spectra (including thermal, shot, and computes efficiently, are useful in system-level analysis as
flicker noises) are handled, and input and output correlations, as
well as individual device contributions, can be calculated. equivalent noise models of RF circuit blocks. They also
HPSD-based analysis is also used to establish the nonintuitive provide an intuitive yet mathematically rigorous visualization
result that bandpass filtering of cyclostationary noise can result of RF noise propagation, which can contribute to design
in stationary noise. insight.
Results from the new method are validated against Monte
Carlo simulations. A large RF integrated circuit (>300 nodes)
The algorithm, an extension of the approach of Ström
driven by a local oscillator (LO) tone and a strong RF signal is and Signell [5], is based on a novel block-matrix relation
analyzed in less than two hours. The analysis predicts correctly between the cyclostationary components of noise within a
that the presence of the RF tone leads to noise folding, affecting circuit. The algorithm can be used to compute the total noise
the circuit’s noise performance significantly. at a specific output, correlations between noise at different
Index Terms— Cyclostationary noise, harmonic balance, har- outputs, and also individual contributions from each noise
monic power spectral density, HPSD, mixer noise, nonlinear noise generator to a specific output. Moreover, bias-dependent white
analysis, RF noise. and colored noise sources (e.g., thermal, shot, and flicker
noises) are treated naturally, even when they are correlated. All
I. INTRODUCTION computations are performed efficiently, i.e., the algorithm can
handle large circuits with many nonlinearities with reasonable

P REDICTING noise in RF circuits is a more complex

task than in linear circuitry, because the former typically
undergo large-signal (quasi-)periodic variations in their op-
time and memory requirements. Efficiency is maintained under
multitone large-signal excitations. A two-tone+noise analysis
of an RF mixer circuit with more than 300 nodes is presented,
eration, unlike the latter. Therefore the statistics of the RF predicting noise folding due to the strong RF input tone.
circuit’s noise also vary (quasi-)periodically in time, leading In a separate application of the block-matrix relation, it is
to important effects such as up- and down-conversion of shown that one-sided (or single-sideband) filtering of cyclo-
noise spectra. Such effects cannot be predicted by traditional stationary noise removes cyclostationary components to leave
SPICE-like noise simulations, which are based on linear-time- stationary noise. This nonintuitive result is confirmed using
invariant (LTI) analysis of stationary noise. In order to address the new algorithm and also through extensive Monte Carlo
such effects, the time-varying nature of noise, as well as the simulations.
variations of the circuit due to its large signal swings, must be The remainder of the paper is organized as follows. In
considered. In this paper, a new RF noise formulation and Section II, the need for cyclostationary analysis is motivated
algorithm is presented which uses cyclostationary stochas- with a simple example, and the concept of harmonic power
tic process and linear-time-varying concepts to capture time spectral densities introduced. The cyclostationary formulation,
variation. block-matrix relation, and algorithm for large circuits are then
Most previous algorithms [1]–[3] that model time-varying presented in Section III. In Section IV, the effect of single-
noise are limited to designs containing relatively few nonlinear sideband filtering of cyclostationary noise is investigated. In
elements, characteristic of microwave circuits. These methods Section V, the new algorithm is verified against Monte Carlo
are impractical for integrated RF circuits where nonlinear de- simulations, and circuit examples presented.
vices are numerous. Recently, an algorithm [4] was proposed
that can analyze large circuits efficiently, but it is limited
to single-tone excitations only. The technique presented in II. CYCLOSTATIONARY NOISE AND HPSD’S
The circuit of Fig. 1 consists of a mixer, followed by a band-
Manuscript received August 5, 1997; revised November 10, 1997.
The authors are with Bell Laboratories, Murray Hill, NJ 07974 USA. pass filter, followed by another mixer. This is a simplification
Publisher Item Identifier S 0018-9200(98)01711-9. of, e.g., the bias-dependent noise generation mechanism in
0018–9200/98$10.00  1998 IEEE
ROYCHOWDHURY et al.: CYCLOSTATIONARY NOISE ANALYSIS OF LARGE RF CIRCUITS WITH MULTITONE EXCITATIONS 325

are termed harmonic autocorrelation functions. The

periodically time-varying power of is its autocorrelation
function evaluated at , i.e., . The quantities
represent the harmonic components of the periodically
varying power. The average power is simply the value of the
dc or stationary component .1 The frequency-domain
representation of the harmonic autocorrelations are termed
harmonic power spectral densities (HPSD’s) of ,
defined as the Fourier transforms

(2)

Equations can be derived that relate the HPSD’s at the

inputs and outputs of various circuit blocks. By solving these
equations, any HPSD’s in the circuit can be determined.
Fig. 1. Mixer-filter-mixer circuit: naı̈ve analysis.
Consider, for example, the circuit in Fig. 1. The input and
output HPSD’s of a perfect cosine mixer with unit amplitude
can be shown [8] to be related by
semiconductor devices [6]. Both mixers multiply their inputs
by a local oscillator (LO) of frequency , i.e., by .
The bandpass filter is centered around and has a bandwidth
of . The circuit itself is noiseless, but the input to
(3)
the first mixer is stationary band-limited noise with two-sided
bandwidth . ( and denoting the input and output, respectively). The
A naı̈ve attempt to determine the output noise power would HPSD relation for a filter with transfer function is [8]
consist of the following analysis, illustrated in Fig. 1. The first
mixer shifts the input noise spectrum by and scales it (4)
by 1/4. The resulting spectrum is multiplied by the squared The HPSD’s of the circuit are illustrated in Fig. 2. Since
magnitude of the filter’s transfer function. Since this spectrum the input noise is stationary, its only nonzero HPSD is
falls within the passband of the filter, it is not modified. Finally, the stationary component , assumed to be unity in the
the second mixer shifts the spectrum again by and scales frequency band , as shown. From (3) applied to
it by 1/4, resulting in the spectrum with three components the first mixer, three nonzero HPSD’s ( and ,
shown in the figure. The total noise power at the output, i.e., shown in the figure) are obtained for . These are generated
the area under the spectrum, is 1/4 that at the input. by shifting the input PSD by and scaling by 1/4; in
This common but simplistic analysis is inconsistent with contrast to the naı̈ve analysis, the stationary HPSD is not
the following alternative argument. Note that the bandpass the only spectrum used to describe the upconverted noise.
filter, which does not modify the spectrum of its input, can From (4), it is seen that the ideal bandpass filter propagates
be ignored. The input then passes through only the two the three HPSD’s of unchanged to . Through (3),
successive mixers, resulting in the output noise voltage the second mixer generates five nonzero HPSD’s, of which
. The output power is only the stationary component is shown in the figure.
This is obtained by scaling and shifting not only the stationary
HPSD of , but also the cyclostationary HPSD’s, which in
fact contribute an extra 1/4 to the lobe centered at zero. The
The average output power consists of only the term,
average output noise (the shaded area under ) equals
since the cosine terms time-average to zero. Hence, the average
3/8 of the input noise.
output power is 3/8 of the input power, 50% more than that
This simple example illustrates how the HPSD approach
predicted by the previous naı̈ve analysis. This is, however, the
can be used to analyze RF noise rigorously yet conveniently.
correct result.
HPSD’s are in fact a powerful tool: incorporating a nonideal
The contradiction between the arguments above underscores
filter in Fig. 1 is simple using (4), and noise propagation
the need for a more rigorous analysis. Modeling circuit noises
through circuits is easy to visualize; this can result in insights
as stochastic processes provides the required generality and
that are otherwise difficult to obtain (e.g., see Section IV). The
rigor. Since the local oscillator is periodic, the processes are
formulation is useful not only for hand calculations and proofs,
cyclostationary [5], [7], i.e., their statistics vary periodically
but also for simulating large circuits, since the HPSD’s of
with time. The autocorrelation function of any cyclostationary
circuit unknowns obey a block-matrix relation. This equation,
process (defined as , with
together with an efficient algorithm to compute it for large
denoting expectation) can be expanded in a Fourier series in
circuits, is described in Section III.
1 Stationary processes are a special case of cyclostationary processes, where
(1) the autocorrelation function (hence the power) is independent of the time t;
 6
it follows that Rz ( ) 0 if i = 0.
326 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 33, NO. 3, MARCH 1998

A. Propagation of Noise Through a Linear

Time-Varying (LTV) System
The input–output relation of the LTV system described by
(6) is

(7)

The objective of this section is to obtain a relation between

the statistics of and if they are stochastic processes. As-
suming that they are nonstationary processes, their covariance
matrices are defined as [5], [7], [9]

(8)

where is or . A straightforward analysis establishes the

following relation between and :

Fig. 2. HPSD’s of mixer-filter-mixer circuit.

(9)
Most nonlinear systems of interest involve periodic wave-
III. EFFICIENT CYCLOSTATIONARY
forms. If , the unperturbed solution of (5), is periodic with
NOISE COMPUTATION ALGORITHM
period , then and of (6) are also -periodic. Hence
The equations of any nonlinear circuit can be expressed in describes a linear periodic time-varying (LPTV)
the form system, and is periodic with respect to displacements
of in both its arguments, i.e.,
(5)
(10)

where are the time-domain circuit variables or unknowns, The periodicity of implies that it can be expanded in a
is a vector of large-signal excitations, and and Fourier series
represent the “resistive” and “dynamic” elements of the circuit,
respectively. The last term represents “small” perturba- (11)
tions to the system, e.g., from noise sources in devices. All
these quantities are vectors of dimension . has dimension
, representing the number of noise sources in the circuit. in the above equation are functions of one variable and will
is an incidence matrix of size which describes how be referred to as the harmonic impulse responses of the LPTV
these noise sources are connected to the circuit. system. Moreover, their Fourier transforms will be denoted by
Since the noise sources are small, their effects can be and referred to as the harmonic transfer functions of the
analyzed by perturbing the noise-free solution of the circuit. LPTV system, i.e.,
Let represent the large signal solution of (5) with set
to zero. Performing a time-varying linearization of (5) about (12)
, the following linearized small-signal differential equation
is obtained: Next, two-dimensional power spectral densities are defined
by taking two-dimensional Fourier transforms of and
(6) [5], [9]

in (6) now represents the small-signal deviations of the

perturbed solution of (5) from the noise-free solution . (13)
and are the derivative matrices of and . By Fourier transforming (9) and using the definitions in
Equation (6) is a linear differential equation with time- (12)–(13), an expression relating and is obtained
varying coefficients. It therefore describes a linear time-
varying (LTV) system with input and output . The
LTV system is characterized completely by its time-varying (14)
impulse response (or kernel) , an matrix. The
dependence of on and will be examined in
Section III-B; the propagation of noise through LTV systems The assumption that both input and output noises are cyclo-
is analyzed next. stationary is now introduced. The cyclostationary assumption
ROYCHOWDHURY et al.: CYCLOSTATIONARY NOISE ANALYSIS OF LARGE RF CIRCUITS WITH MULTITONE EXCITATIONS 327

implies that and do not change if is added to both obtained for the two-dimensional power spectral densities
arguments, i.e., and [5]:
(15)
(18)
Hence both can be expressed as Fourier series
Using (18), the relation between the two-dimensional power
(16) spectral densities [(14)] is rewritten in terms of the (one-
dimensional) harmonic PSD’s and
and are functions of one variable and will be
referred to as the harmonic covariances of the output and input
noise, respectively. Their (one-dimensional) Fourier trans- (19)
forms will be denoted by and and referred to as
harmonic PSD’s or HPSD’s, i.e.,
Equation (19), relating the harmonic PSD’s of the input and
(17) output noise, is an extension to cyclostationary noise inputs
of a similar equation by Ström and Signell [5]. An interesting
The harmonic covariances and PSD’s have simple physical and useful observation about (19) is that the output harmonic
interpretations. represents the time-varying power appears only in the last term . This suggests that
of the cyclostationary noise; hence by (17), (the (19) can be written in block matrix form. It can be verified by
harmonic covariances evaluated at zero) represent the Fourier direct multiplication that (19) is equivalent to the following
components of the periodically varying noise power. In partic- block matrix equation:
ular, is the average value, or stationary component, (20)
of the power. From the definition of the harmonic PSD’s, it
follows that the harmonic covariances evaluated at zero are where denotes the Hermitian of . and are
equal to the corresponding harmonic PSD’s integrated over block matrices with an infinite number of blocks, shown at
the entire frequency axis. Hence integrated equals the bottom of the page in (21)–(23) (denoting by
the stationary component of the output noise power. and for conciseness).
will be therefore be termed stationary PSD’s. Equation (20) expresses the relation between the output
When the -periodic assumption of (15) and the definitions and input harmonic PSD’s compactly using block matrices.
of (16) and (17) are applied to (13), the following form is Note from (22) that the output harmonic PSD’s evaluated

.. .. .. .. ..
. . . . .

(21)

.. .. .. .. ..
. . . . .
.. .. .. .. ..
. . . . .

(22)

.. .. .. .. ..
. . . . .
.. .. .. .. ..
. . . . .

(23)

.. .. .. .. ..
. . . . .
328 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 33, NO. 3, MARCH 1998

at are given by the central block-row of . The ..

.
HPSD’s of the self- and cross-powers of the th output are
available in the th row of this block. Denote the transpose of (30)
this row by ; this is obtained by transposing (20)
and postmultiplying by a unit block-vector followed by ..
.
the th unit vector ..
.
(24)

where
(31)
..
. st entry
..
. ..
.. .
. (25)
th block th entry is known as the conversion matrix [10] of the circuit;
..
. is the Jacobian matrix of the harmonic balance equations
.. th entry at the circuit’s steady state .
.
For numerical computation, the infinite block matrices in
where represents the identity matrix. Note that (21)–(23) and (28)–(31) are truncated to a finite number of
is a vector. Hence the computation of blocks . is the largest positive harmonic
in (24) can be performed by matrix-vector products with the considered. For the purposes of the analysis, it is assumed
block matrices and . Despite the fact that these that no significant harmonic PSD of degree greater than
matrices are, in general, dense, products with them can be exists for the input noise or the output noise . Since
performed efficiently, as discussed next in Sections III-B and the energy content of the th harmonic is always a diminishing
III-C. function of in practical RF circuits, a value for can always
be found satisfying this assumption.
B. Fast Application of and With this assumption, it can be shown that the Toeplitz
Exploiting Harmonic Balance block structure in the above matrices can be approximated
by circulant block structure without loss of accuracy in the
To apply and efficiently to a vector, it is necessary to
matrix-vector product. For example, truncated to
represent in terms of and [refer to (6)]. Since
and are -periodic, they are expanded in Fourier blocks can be approximated by , given by
series

(26)
(32)
The Fourier coefficients and will be referred to as the
harmonics of and , respectively. It can be shown
[10] that can be expressed in terms of these harmonics as
(27) Note that the fourth, fifth, and sixth sub- and super-diagonals
of differ from those of truncated to seven blocks. Matrix-
where
.. .. .. .. .. vector products with and the truncated , however, produce
. . . . . identical results up to the first harmonic location if the vector
being multiplied contains no significant components in the
second and third harmonic locations.
(28) The utility of the circulant approximation is that it enables
and to be decomposed into products of sparse block-
diagonal matrices, permutations, and Fourier transform (DFT)
.. .. .. .. ..
. . . . . matrices [11]–[13]. This enables matrix-vector products with
.. .. .. .. .. and to be performed as a sequence of products with sparse
. . . . .
block-diagonal matrices ( operations), permutations
(no cost), and Fourier transforms ( operations);
(29) hence the overall computation is . Further, since
only the sparse block-diagonal matrices need to be stored,
the memory requirement is . Note that is a
.. .. .. .. .. diagonal matrix with a priori known entries , hence its
. . . . . application to a vector is in computational cost, with
ROYCHOWDHURY et al.: CYCLOSTATIONARY NOISE ANALYSIS OF LARGE RF CIRCUITS WITH MULTITONE EXCITATIONS 329

no memory required for its storage. The net effect of the circuits operating in time-varying steady state, unfortunately,
circulant approximation, therefore, is that can be applied there are as yet no stochastic models for the nonstationary
to a vector in computation and memory. noise generation process that are well established. Neverthe-
From (27), it follows that to obtain the required matrix- less, there is general consensus that for white processes like
vector product with , matrix-vector products with and shot and thermal noise, the time variation is generated by
are necessary. Since is a sparse block-diagonal modulating stationary white noise by the (deterministic) time-
matrix with identical blocks (the noise source incidence varying large-signal steady state. For the diode shot noise
matrix), it can be applied in time and memory. example above, (34) generalizes to
Iterative linear solvers [14], [15] can obtain the matrix-vector
product with using only matrix-vector products with (35)
. The use of iterative linear techniques, together with the
decomposition of allowing its application in where is a time-varying waveform. Arguments sup-
time, is the key to the fast harmonic balance algorithms of porting this deterministic modulation model are based on the
Rösch [12], [13], [16] and Melville et al. [11]. With suitable short-term nature of the autocorrelations of thermal and shot
preconditioning included in the iterative solution, the number noise; see, e.g., [3], [18], and [19].
of -vector products required to compute a -vector product For noise with long-term correlations (notably flicker
is small and approximately independent of the size of . Hence noise), there is a general belief that the above deterministic-
the -vector product can be computed in approximately modulation-of-stationary-noise model is inadequate (e.g.,
time and memory, leading to the same [20]). The physical processes responsible for generating long-
computation and memory requirements for the desired product term noise correlations, it is argued, are themselves modified
with . by the large-signal waveforms which change on a relatively
From Section III-A, products are required with and faster scale. Unfortunately, neither theoretical analyses nor
for cyclostationary noise computation. Application of experimental data are available at this time, to the authors’
is carried out using the same decomposition and iterative knowledge, to aid in formulating a generation mechanism
linear methods as for , but using transposes of the matrices for such noise. In the absence of an established alternative,
involved. The product with is carried out using the relation Demir [18] has used the modulated stationary noise model
. for analyzing nonstationary flicker noise, and this approach
also appears common among designers of RF circuits. The
C. Fast Application of modulated stationary noise model is therefore reluctantly
adopted in this work for all cyclostationary noise generators.
The principal sources of noise in circuits are thermal, shot, Under this noise generation model, the noise input in
and flicker noises from devices. When the linearized (6) can be expressed as
small-signal circuit [(6)] is time-invariant (i.e., the circuit
is in dc steady state), these noise sources are stationary (36)
stochastic processes with known power spectral densities.
Thermal and shot noises are white, i.e., their PSD values are where is an -dimensional vector of stationary noise
constant, independent of frequency; flicker noise PSD’s exhibit sources and is an diagonal matrix of -periodic
a variation with frequency. The expressions for the power deterministic modulations.
spectral densities of these noise sources (see, e.g., Van der Ziel Equation (20) can be used to analyze the relation between
[17]) typically involve some component of the dc solution; statistics of and by recognizing that (36) represents
for example, the PSD of the shot noise current across an LTV system with input and output . The time-
a diode’s p-n junction is proportional to the dc current varying impulse response of the LTV system is
through the junction, i.e.,
(33)

where is the (stationary) PSD of the shot noise and (37)

is the electronic charge. where denote the Fourier coefficients of the periodic
From the viewpoint of second-order statistics, the diode’s modulation . The harmonic transfer functions are
shot noise is equivalent to the hypothetical process generated independent of and simply equal to . Equation (20)
by multiplying a white noise process of PSD value applied to this LTV system results in the following block-
by a constant factor of matrix relation between the harmonic PSD’s of and :

(34) (38)

For this reason, shot noise is often said, in a loose sense, to where represents the block Toeplitz matrix of the
be proportional to . harmonic PSD’s of the stationary noise sources. Since the
For circuits operating in dc steady state, expressions for sources are stationary, all their harmonic PSD’s are zero except
PSD’s of stationary noise generators are well established from for the stationary PSD ; hence is block di-
theoretical considerations and/or through measurement. For agonal with diagonal entries
330 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 33, NO. 3, MARCH 1998

. in (38)
is block Toeplitz with in the diagonals
.. .. .. .. ..
. . . . .

(39)

.. .. .. .. ..
. . . . .

Using (38), the product of with a vector can be

performed through matrix-vector products with the matri-
ces and . Products with the block-Toeplitz
matrices and can be performed in
time and memory, approximating by a block-
circulant matrix and applying the same decomposition as for
in Section III-B. Application of the block-diagonal matrix
is equivalent to matrix vector products with
. If the device noise generators are uncorrelated,
is diagonal; if correlations exist, they are usually
between small groups of noise generators, hence Fig. 3. H (!) overlaid with H (! + i!0 ) for i = 2; 1; 01 and 02.
is sparse. In either case, each product with is
in computation with no storage required. Hence matrix-
vector products with time and memory.
The overall matrix-vector product with can therefore be
performed in time and memory.
It should be noted that the noise modulation can be
absorbed into the circuit equations [(5)]. The noise inputs
to the circuit can then be assumed to be stationary without loss
of generality. This procedure, however, increases the size of
the harmonic balance system for obtaining the steady state .
To avoid this and to separate the implementation of the noise
algorithm from the harmonic balance steady state algorithm,
the formulation of this section is preferred.

IV. BANDPASS FILTERING OF CYCLOSTATIONARY NOISE

Fig. 4. One-sided H (!) overlaid with H (! + i!0 ) for i = 2 and 02.
Ström and Signell [5] have shown that low-pass filtering
of cyclostationary noise results in stationary noise if the
bandwidth of the low-pass filter is less than half the frequency Note that the th harmonic PSD of the output is determined
of cyclostationarity . This result has been used by Hull completely by the corresponding harmonic PSD of the input,
and Meyer [21] to simplify their analysis. In this section, the shaped by the product of the filter function with a shifted
effect of LTI bandpass filtering on cyclostationary noise is version of itself . For the scalar input–output case
considered. It is shown that if cyclostationary noise is passed under consideration, the relation simplifies to
through a one-sided (i.e., single-sideband) bandpass filter of
bandwidth less than , the output noise is stationary. (41)
This result is obtained using a simple visualization of the
propagation of harmonic PSD’s. Since the magnitude of for a real filter is symmetric
Denote the input noise to a bandpass filter by and about zero, has the same magnitude characteristic as
the output noise by . Assume that the input is . By overlaying the magnitudes of and
cyclostationary with period . Denote the transfer for different values of (illustrated in Fig. 3), it can be
function of the bandpass filter by . The relationship seen that the product is nonzero only for
between the harmonic PSD’s of and , derived from and if the bandwidth of is less than .
(24) by using the fact that is block diagonal for an LTI For all other values of , there is no frequency at which
network, is and are both nonzero, hence their product is
identically zero.
(40) This immediately implies the following.
ROYCHOWDHURY et al.: CYCLOSTATIONARY NOISE ANALYSIS OF LARGE RF CIRCUITS WITH MULTITONE EXCITATIONS 331

Fig. 5. Mixer and bandpass filter.

Fig. 6. Bandpass filter characteristic. Fig. 7. yy (f ) with f0 = 1:592 MHz (double-sideband filtering).
S

Result 1: Bandpass filtering with bandwidth less than circuit, consisting of an I-channel buffer and mixer, driven by
eliminates all harmonic PSD’s except the stationary and sec- two strong tones (a signal and an LO).
ond harmonic PSD’s.
Moreover, if the bandpass filter is one-sided with respect A. Mixer and Bandpass Filter
to , then the product is identically zero Motivated by the result of Section IV, a mixer and bandpass
also for and , as illustrated in Fig. 4. In this case, the filter circuit (Fig. 5) is analyzed for cyclostationary noise
bandwidth of the filter can be greater than but should be propagation. The mixer is an ideal multiplier that modulates
less than . The only nonzero PSD of the output is then the the incoming stationary noise with a deterministic LO os-
stationary PSD. This implies Result 2. cillator signal . The filter has a high- bandpass
Result 2: One-sided (or single-sideband) bandpass filtering characteristic (illustrated in Fig. 6) with a center frequency of
(with bandwidth less than ) of cyclostationary noise results approximately 1.592 MHz and a bandwidth of about 50 kHz.
in stationary output noise. The stationary input noise is bandlimited with double-sided
bandwidth of about 200 kHz.
V. RESULTS Two simulations are carried out, with set to 1.592 MHz
The fast cyclostationary noise algorithm of Section III has and 1.5 MHz, respectively. In the first situation, the bandpass
been prototyped in a Bell Labs internal simulator. In this filter is centered at the LO frequency; in the second, the
section, the algorithm and the single-sideband-filtering results filter characteristic is offset to the right of the LO frequency,
of Section IV are first verified against Monte Carlo noise strongly attenuating the lower sideband with respect to
simulations with 60 000 sample waveforms, to an accuracy of while passing the upper sideband. Harmonic PSD’s at all nodes
within 2%. Noise analysis results from two circuits are then in the circuit were computed over frequencies from 1 MHz to
presented—a mixer excited by a single LO tone, and a large 2 MHz.
332 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 33, NO. 3, MARCH 1998

Fig. 8. yy
S (f ) with f0 = 1:592 MHz (double-sideband filtering). Fig. 10. S yy (f ) with f0 = 1:5 MHz (single-sideband filtering).

Fig. 9. S yy (f ) with f0 = 1:5 MHz (single-sideband filtering). Fig. 11. Time-varying filtered noise power from Monte Carlo:
f0 = 1:592 MHz (double sideband).

Using the results of Section III-A, it can be shown that only

the stationary and second harmonic PSD’s of the mixer output
are nonzero, related to the PSD of the stationary input by have the same magnitude, hence there is a large cyclostationary
component in the noise. The same PSD’s for the MHz
case (single-sideband filtering) are shown in Figs. 9 and 10.
The second harmonic PSD can be seen to be about two orders
of magnitude smaller than the stationary PSD. Hence the
filtered noise is virtually stationary, as predicted in Section IV.
The stationary and second harmonic PSD’s of the filter The second harmonic PSD is not identically zero because the
output for MHz (double-sideband filtering) nonideal single-sideband filter does not perfectly eliminate the
are shown in Figs. 7 and 8. It can be seen that both PSD’s lower sideband.
ROYCHOWDHURY et al.: CYCLOSTATIONARY NOISE ANALYSIS OF LARGE RF CIRCUITS WITH MULTITONE EXCITATIONS 333

cyclostationarity of the noise in the double-sideband case can

be seen from the variation of the power between zero and its
maximum value of about 0.0022. In the single-sideband case,
the power approaches a steady value of about 130 10 ,
with a cyclostationary variation of about 10%. Accounting
for the normalization of the input PSD, these values are in
excellent agreement with the total integrated noise of
and (Figs. 9 and 10); Monte Carlo simulation results
are within 2% of the results produced by the new algorithm.

B. Mixer Analysis
The mixer in Fig. 13 was analyzed for cyclostationary noise
to investigate the effect of large-signal LO variations on the
output noise. The LO signal of amplitude 1.5 V is applied at
the base of the first transistor, as shown. The RF input signal
is applied through the current source, which is held at a dc
value of 2 mA (i.e., no RF signal) for the noise analysis.
Two simulations were performed: a stationary analysis with
no LO present to obtain the noise of the quiescent circuit, and
a cyclostationary analysis with the LO amplitude at 1.5 V. The
Fig. 12. Time-varying filtered noise power from Monte Carlo:
former simulation took a few seconds and the latter (with 25
f0 = 1:5 MHz (single sideband). large-signal harmonics) 40 seconds per frequency point. The
stationary PSD is shown in Fig. 14, and some nonstationary
HPSD’s in Fig. 15.
From Fig. 14, it can be seen that the presence of a large
LO signal reduces the average noise power at the output. This
is a known property of switching mixers. Fig. 15 shows the
first six harmonic PSD’s of the noise at the output when the
LO is 1.5 V.
From a knowledge of the HPSD’s, it is possible to create
system-level macromodels for functional blocks like the mixer.
All the noise of the circuit can then be concentrated in an
equivalent noise source with the same HPSD’s. While only
the stationary PSD determines the average noise power, the
nonstationary HPSD’s must be included because they can
contribute to the stationary component of some other block,
as discussed in Section II.

C. I-Channel Buffer and Mixer Circuit

The next example is a portion of the W2013 RFIC, con-
Fig. 13. Mixer cell. sisting of an I-channel buffer feeding a mixer. The circuit
consisted of about 360 nodes and was excited by two tones—a
The above results were also verified by simulations using local oscillator at 178 MHz driving the mixer and a strong RF
the Monte Carlo method. The nonlinear differential equations signal tone at 80 kHz feeding into the I-channel buffer. Two
of the circuit in Fig. 5 were solved numerically with 60 000 noise analyzes were performed. The first analysis included
samples of the input noise [22]. The input noise PSD both LO and RF tones (sometimes called a three-tone noise
was normalized to one to avoid corruption of the results analysis). The circuit was also analyzed with only the LO tone
by numerical noise generated during differential equation to determine if the RF signal affects the noise significantly.
solution. The 60 000 samples of the mixed and filtered noise The two-tone noise simulation, using a total of 525 large-
were squared and averaged, on a per-timepoint basis, to signal mix components, required 300 MB of memory and
obtain the noise power at the output as a function of time. for each frequency point, took 40 min on an SGI machine
The variation of noise power with time is shown in Figs. 11 (200-MHz R10000 CPU). The one-tone noise simulation,
and 12 for the double-sideband and single-sideband cases. using 45 harmonics, needed 70 MB of memory and took 2 min
When analyzed in the time domain, the circuit requires some per point.
time to reach large-signal steady state, hence the steady state The stationary PSD’s of the mixer output noise for the
noise power is approached toward ms; in contrast, two simulations are shown in Fig. 16. It can be seen that
harmonic balance calculates this steady state directly. The the presence of the large RF signal increases the noise by
334 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 33, NO. 3, MARCH 1998

Fig. 14. Mixer: stationary PSD at output.

Fig. 15. Mixer: harmonic PSD’s at output.

ROYCHOWDHURY et al.: CYCLOSTATIONARY NOISE ANALYSIS OF LARGE RF CIRCUITS WITH MULTITONE EXCITATIONS 335

[2] A. Demir, E. Liu, and A. Sangiovanni-Vincentelli, “Time-domain non

Monte-Carlo noise simulation for nonlinear dynamic circuits with ar-
bitrary excitations,” IEEE Trans. Computer-Aided Design, vol. 15, pp.
493–505, May 1996.
[3] M. Okumura, H. Tanimoto, T. Itakura, and T. Sugawara, “Numerical
noise analysis for nonlinear circuits with a periodic large signal ex-
citation including cyclostationary noise sources,” IEEE Trans. Circuits
Syst.–I, vol. 40, pp. 581–590, Sept. 1993.
[4] R. Telichevesky, K. Kundert, and J. White, “Efficient AC and noise
analysis of two-tone RF circuits,” in Proc. IEEE DAC, 1996, pp.
292–297.
[5] T. Ström and S. Signell, “Analysis of periodically switched linear
circuits,” IEEE Trans. Circuits Syst., vol. CAS-24, pp. 531–541, Oct.
1977.
[6] A. R. Kerr, “Noise and loss in balanced and subharmonically pumped
mixers: Part 1—Theory,” IEEE Trans. Microwave Theory Tech., vol.
MTT-27, pp. 938–943, Dec. 1979.
Fig. 16. Stationary PSD’s for the I-Q mixer/buffer circuit. [7] W. Gardner, Introduction to Random Processes. New York: McGraw-
Hill, 1986.
[8] J. Roychowdhury and P. Feldmann, “A new linear-time harmonic
balance algorithm for cyclostationary noise analysis in RF circuits,”
in Proc. ASP-DAC, 1997, pp. 483–492.
about 1/3. This is due to noise folding, the result of devices [9] A. Papoulis, Probability, Random Variables and Stochastic Processes.
being driven into nonlinear regions by the strong RF input New York: McGraw-Hill Series in System Science, 1965.
tone. This effect is difficult to predict with the technique of, [10] S. A. Haas, Nonlinear Microwave Circuits. Norwood, MA: Artech
House, 1988.
e.g., [4]. The peaks in the two waveforms, located at the LO [11] R. C. Melville, P. Feldmann, and J. Roychowdhury, “Efficient multitone
frequency, are due to up- and down-conversion of noise from distortion analysis of analog integrated circuits,” in Proc. IEEE CICC,
May 1995, pp. 241–244.
other frequencies. [12] M. Rösch, “Schnell simulation des stationären Verhaltens nichtlinearer
Schaltungen,” Ph.D. dissertation, Technischen Universität München,
1992.
VI. CONCLUSION [13] H. G. Brachtendorf, G. Welsch, R. Laur, and A. Bunse-Gerstner, “Nu-
merical steady state analysis of electronic circuits driven by multitone
signals,” Elect. Eng., vol. 79, pp. 103–112, 1996.
A frequency-domain formulation and algorithm has been [14] R. Freund, G. H. Golub, and N. M. Nachtigal, “Iterative solution of
presented for computing noise in nonlinear circuits. The linear systems,” Acta Numerica, pp. 57–100, 1991.
[15] Y. Saad, Iterative Methods for Sparse Linear Systems. Boston, MA:
method uses cyclostationary components and harmonic PSD’s PWS, 1996.
in its formulation to capture time-varying noise statistics. A [16] M. Rösch and K. J. Antreich, “Schnell stationäre simulation nichtlinearer
block-structured matrix equation for the output noise statistics Schaltungen im Frequenzbereich,” AEÜ, vol. 46, no. 3, pp. 168–172,
1992.
is the central result enabling fast computation. The algorithm [17] A. Van der Ziel, Noise in Solid State Devices and Circuits. New York:
is efficient for large circuits with several large tones and can Wiley, 1986.
[18] A. Demir, “Time domain non Monte Carlo noise simulation for non-
generate information useful for noise macromodels. linear dynamic circuits with arbitrary excitations,” Tech. Rep. Memo.
The new formulation has been used to prove that one-sided UCB/ERL M94/39, EECS Dept., Univ. Calif. Berkeley, 1994.
bandpass filtering of cyclostationary noise produces stationary [19] V. Rizzoli and A. Neri, “State of the art and present trends in nonlinear
microwave CAD techniques,” IEEE Trans. Microwave Theory Tech.,
noise. This extends a previously known result for low-pass vol. 32, pp. 343–365, Feb. 1988.
filtering. [20] M. Banu, V. Gopinathan, R. Meyer, and Y. Tsividis, Personal commu-
The algorithm been verified against Monte Carlo simu- nications, 1996.
[21] C. Hull and R. G. Meyer, “A systematic approach to the analysis of
lations. Results from a mixer cell and a large I-channel noise in mixers,” IEEE Trans. Circuits Syst.–I, vol. 40, pp. 909–919,
buffer and mixer RF integrated circuit have been presented, Dec. 1993.
predicting the fact that the presence of multiple tones can [22] J. Roychowdhury and P. Feldmann, “Generating stationary noise for
circuit noise simulation,” Bell Laboratories Tech. Memo., Oct. 1996.
significantly affect the noise performance of a circuit.

ACKNOWLEDGMENT

The authors would like to thank A. Demir, M. Banu, L.

Toth, A. Dec, K. Suyama, V. Gopinathan, and B. Meyer for
useful discussions regarding the RF noise problem. Thanks
are due to A. Mehrotra for help with LaTeX. The efficient Jaijeet Roychowdhury received the B.Tech. degree
from the Indian Institute of Technology, Kanpur,
harmonic balance algorithm used in this work was developed in 1987 and the M.S. and Ph.D. degrees from the
with their colleague R. Melville. University of California, Berkeley, in 1989 and
1993, all in electrical engineering.
From 1993 to 1995, he was with the CAD Lab-
oratory of AT&T Bell Labs, Allentown, PA. Since
REFERENCES 1995, he has been with the Communications Re-
search Division of Bell Laboratories, Murray Hill,
[1] V. Rizzoli, F. Mastri, and D. Masotti, “General noise analysis of nonlin- NJ. His main research interest is the analysis of
ear microwave circuits by the piecewise harmonic-balance technique,” dynamical systems, particularly those relevant from
IEEE Trans. Microwave Theory Tech., vol. 42, pp. 807–819, May 1994. circuits and communication systems.
336 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 33, NO. 3, MARCH 1998

David Long received the B.S. degree from the Peter Feldmann (S’87–M’91–SM’96) was born in
California Institute of Technology, Pasadena, in Timişoara, Romania. He received the B.Sc. degree
1987 and the Ph.D. degree in computer science summa cum laude in computer engineering in 1983
from Carnegie Mellon University, Pittsburgh, PA, and the M.Sc. degree in electrical engineering in
in 1993. 1987, both from the Technion, Israel, and the Ph.D.
Since then he has been a Member of the Technical degree in 1991 from Carnegie Mellon University,
Staff at Bell Laboratories, Murray Hill, NJ. His Pittsburgh, PA.
research interests include circuit simulation, formal From 1985 through 1987 he worked for Zoran
verification, and computational electromagnetics. Microelectronics, Haifa, Israel, on the design of dig-
ital signal processors. Currently, he is Distinguished
Member of Technical Staff at Bell Laboratories,
Murray Hill, NJ. His main research interests are simulation and analysis of
circuits and communication systems.