IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 52, NO.

6, JUNE 2005 1211

Description of Limit Cycles in

Sigma–Delta Modulators
Derk Reefman, Josh Reiss, Erwin Janssen, and Mark Sandler

Abstract—A mathematical framework, based on state-space

modeling, for the description of limit cycles (LCs) of 1-bit
sigma–delta modulators (SDMs) is presented. It is proved that
periodicity in bit output pattern of the SDM implies a periodic
orbit in state-space variables. While the state-space description is
generally applicable for periodic inputs, the focus is on dc inputs,
since this represents the most relevant practical condition. An
outcome of the analysis is that, in general, for an th-order SDM,
at least 1 initial conditions need to be fixed in order to have
LC behavior. Expressions for the minimum disturbance of the
Fig. 1. Illustration of the definitions used in the paper to distinguish a LC (left)
input or initial conditions that is needed to break up a LC are also
from an idle tone (right). An LC consists of a finite number of discrete peaks
presented. Special focus is given to the case where the disturbance in the frequency spectrum; an idle tone is a peak in the frequency spectrum, but
takes the form of “dithering the quantizer”, and it is shown that superposed on a noise background.
this form of dither is a suboptimal approach to remove LCs. The
stability of LCs is determined, and it is demonstrated that a res-
onator section, as often employed to increase the dynamic range of important progress in the description of SDMs is reported in
an SDM, has an adverse effect on LC behavior in that it stabilizes the work of Risbo [3] and Hein and Zakhor [4], while a very
LCs. Furthermore, the experimental observation that high order useful linearization technique is described in [5] and further
SDMs are less susceptible to LCs is underpinned. Finally, some
elaborated on by Magrath [6]. Yet in all these developments,
examples are provided which illustrate the theoretical results.
there is no unified description of SDMs. Instead, several models
Index Terms—Analog–digital (A/D) conversion, digital–analog are provided, each of which describes some aspects of an SDM
(D/A) conversion, limit cycles (LCs), nonlinear circuits, nonlinear
systems, sigma–delta modulation. to a certain accuracy. Though it is not the intention of this paper
to give a comprehensive theory of sigma–delta modulation in
general, it does give an exact mathematical framework. Within
I. INTRODUCTION this framework, some aspects of sigma–delta modulation can
be quantitatively understood. In this paper, the focus is on
S IGMA–DELTA modulation is an increasingly popular
technique for coding data streams [1]. The technique
has provided powerful means for converting analog to dig-
the characterization of limit cycles (LCs). The analysis will
be restricted to the class of “feedforward SDMs” (also called
ital signals and vice versa with low circuit complexity and “interpolative SDMs”) [1], which represents a class of often
large robustness against circuit imperfections. As a result used SDM topologies. The analysis can be easily adapted to
of this, 1-bit sigma–delta based analog-to-digital (A/D) and deal with other topologies as well [7].
digital-to-analog (D/A) converters are widely used in audio We use the definition of an LC which is customary in the
applications, such as cellular phone technology and high-end world of sigma–delta design engineers. An LC is a sequence
stereo systems. In particular, it has seen a further boost in of output bits, which repeats itself indefinitely. As such, we
interest due to the introduction of super audio compact disk want to contrast this definition with that of an “idle tone,” which
(SA-CD). SA-CD is based on a digital format coined DSD, is a discrete peak in the frequency spectrum of the output of an
which is a 1-bit coded representation of the audio stream with SDM, but superposed on a background of noise (see Fig. 1).
a sample rate of 2.8 MHz. In the SA-CD standard, no reference Hence, in that case, there is no unique series of bits which
is made to the technique which is employed to create such repeats itself.
a 1-bit stream, yet sigma–delta modulation is clearly one of Fundamental work on LCs in SDMs has usually been con-
the techniques that is capable of creating a high-quality 1-bit strained to low order SDMs [8]–[10], and hence is of little
stream. Sigma–delta modulation, originally conceived by de practical value to engineers who use high order noise shaping
Jager [2], is a well-established technique. However, theoretical techniques. The basis for the approach used in this paper
understanding of the concept is very limited [1]. The most has been provided in previous work, most notably in [4] and
[11]. An approach which bears some resemblance to the one
presented in the current paper is discussed in [12]. In [12],
Manuscript received June 12, 2004; revised September 17, 2004. This paper
was recommended by Associate Editor M. Chakraborty. however, the relative frequency content of possible LCs in
D. Reefman and E. Janssen are with the Philips Research, 5656 AA Eind- some specific SDMs is discussed. We derive results for a
hoven, The Netherlands. general SDM and focus on the character of the LCs, and
J. Reiss and M. Sandler are with the Queen Mary University of London,
London E14NS, U.K. ([email protected]). their stability in particular. An important assumption in this
Digital Object Identifier 10.1109/TCSI.2005.849134 earlier work is that a periodic bit output pattern implies a
1057-7122/$20.00 © 2005 IEEE
1212 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 52, NO. 6, JUNE 2005

where is the output bit at clock cycle , is the quan-

tizer input signal, and are the integrator outputs, called state
variables. The are the feedforward coefficients, and the the
resonator coefficients determining the positions of the poles in
the loopfilter. The propagation of the states can be written in
matrix notation as

(2)

where is called the transition matrix of the SDM of

order ; describes how the input and
Fig. 2. States in fifth-order SDM.
feedback are distributed. The power of the state-space descrip-
tion is that it allows us to create a very compact description of
periodic orbit in state-space variables. This assumption is not the propagation of the SDM from time to time , as
obvious, but turns out to be true as will be proven in this paper. repeated application of (2) to leads to
Based on a state-space description, we will present an exact
description of LCs in SDMs. While drawing on some known
(3)
results from linear algebra, some remarkable results for LCs
in SDMs are obtained, such as the persistance of LCs while
dithering the SDM. Although in its pure definition, a LC is From the above equation, it is clear that all useful information
a periodic pattern of infinite duration, in practical situations is embedded in the second term of the right-hand-side (RHS) of
finite duration periodic sequences can be equally annoying. (3); the first term carries no input signal information.
The finite duration patterns touch upon the important subject
of stability of a LC. How much time it takes until a small B. General Formulation of LC Conditions
perturbation moves a LC out of its periodic pattern. In the introduction, the following definition of a LC was pro-
The paper is organized as follows. In Section II, the math- vided.
ematical framework, based on a state-space description of the An LC is a sequence of output bits, which repeats itself
SDM, is presented. All the following chapters are based on this indefinitely.
formulation. In Section III, this formulation is applied to prac- In dynamical systems theory, an LC of period exists if, for
tical SDM designs. Some basic quantitative criteria, necessary initial conditions , is the smallest positive integer such
for determining the stability of an LC, are developed. In Sec- that
tion IV, a stability analysis of LCs is presented. The analysis
provides insight into how long it takes before a perturbed SDM (4)
steps out of a LC. In Section V, the concepts of the foregoing
sections will be used to obtain numerical results. Finally, in Sec- for all greater than or equal to zero. The state-space variables
tion VI, conclusions will be presented. then describe a periodic orbit in state space. However, from a
practical point of view, an LC represents periodic behavior in
II. MATHEMATICAL BACKGROUND the output . Henceforth, we will refer to LC when periodic
behavior in the output is meant, and “periodic orbit in state
A. State-Space Description space” when periodicity in the state variables is meant. In Ap-
Though state-space descriptions of discrete-time processes pendix A, it is proven that, under reasonable assumptions, pe-
are well established [13], in this section certain aspects are re- riodicity in implies that a periodic orbit in state space exists.
viewed in order to present the paper in a self-contained way. The However, such equivalence is by no means always true. Thus,
state-space description of the SDM in Fig. 2 will be examined as under the assumptions stated in Appendix A, there is a strict set
an illustration of a feedforward topology. This figure displays a of necessary (but not sufficient!) equalities that need to hold for
feedforward SDM of order with two resonator sections the initial states if periodic output is to be sustained
and the optional inclusion of dither just prior to quantization.
This represents a typical modulator design, which is often used
in practical designs [14]. The state-space description of other
classes of modulators closely resembles the description that fol- (5)
lows below [15]. One can easily read that, for the SDM depicted where has been introduced to avoid cumbersome
there notation. Formally

(6)

From (5), one can obtain a unique value for the initial state
(1) if, and only if, the inverse of the matrix exists. This will
REEFMAN et al.: DESCRIPTION OF LCS IN SDMs 1213

be elaborated in Section III; for now, it is assumed that a solution important case, and as its effectiveness is strongly related to the
or solution space to (6) exists. LC conditions, its discussion is included in Section III.C.
So far, the appearance of the LC has not been specified, ex-
cept that it is of period . If a LC is now defined as a specific A. SDMs With DC Poles
sequence ( ), then for each In the case that the SDM has at least one of its poles at dc, the
matrix is singular, and hence not invertible. To solve
(7) (5) for that case, one may use the singular-value decomposition
(SVD) [16] of
which is a test that has to be passed if a LC of the specified se-
quence ( ) exists. The inaccuracy made (10)
in (7) is that the possibility that has been left out.
As this equality occurs with probability zero over the conti- where is a diagonal matrix whose elements
nously variable value of , this should not pose much of are the singular values of . The matrices
a problem. Thus, there is a set of equalities, (5), and a set of in- and are the left and right singular vectors, re-
equalities, (7), that need to be fullfilled in order to have a valid spectively. Because both and are unitary, we also have
LC. Substitution of (3) in (7) gives . When the SDM is not reducible, ex-
actly one of the singular values will be zero as a result of the
fact that the loop filter displays a pole at dc. When the singular
values are ordered in descending fashion, this singular value will
be . This has the interesting consequence, that the last
(8)
column of is a nonrelevant direction, since it is always mul-
tiplied by . This last column of will be denoted
which is equivalent to (the so-called null space of : ). Now,
if a single solution is known (say, ) to (5), any solution
can be expressed as
(9)
by defining . (11)
Hence, one needs to simultaneously solve (5) and (9) in order
to have a valid LC; in the next section more specific solutions In other words, the complete set of solutions to (5) is a line.
will be derived for various SDM topologies. Thus, for an order SDM, at least (initial) conditions
need to be fulfilled in order to have a LC.
In addition, the SVD is helpful in obtaining an initial solution
III. LIMIT CYCLE CONDITIONS FOR SPECIFIC
to (5). Similar to (10),(11)
SDM ARCHITECTURES
In order to quantify the importance of any disturbance of a (12)
LC, one first needs to solve (6). However, in the previous sec-
tion, the remark has been made that the matrix may where is the null space of . Therefore, (12) is
not be invertible. This observation carries significant practical equivalent to
relevance. The poles of the loop filter of an SDM are given by
the eigenvalues of the transition matrix . Each pole can (13)
be written as , where is the pole frequency [4].
Hence, for a classical SDM which has all its loopfilter poles at Multiplying both sides of (5) with one obtains
dc, all eigenvalues of will be one, as a result of which the
inverse of (5) does not exist – hence, there is no unique so- (14)
lution to . On the other hand, if one has an SDM of even
order with resonator sections, all loopfilter poles will stating a necessary condition for the exis-
occur for frequencies other than dc. As a result, there exists one tance of a solution to (5). For the type of SDMs that are investi-
and one only initial state that results in a specific LC. Most gated in this section, with a pole at dc (and thus infinite gain for
often, SDMs have at least a single zero at dc to avoid dc drift. In dc) this condition is equivalent to the intuitively obvious con-
the following, we will distinguish between two main categories dition that the average input should equal the average output of
of SDMs. Those with and without poles at dc. The SDMs with the SDM
poles at dc will be further subdivided in two categories. Those
with poles at dc for the last two integrator sections; and those (15)
with poles away from dc for the last two integrator sections.
A special case of LC break up is due to dithering the SDM.
Typically, dithering is achieved by adding a random number to When the SDM input is a constant dc value, and the
the input of the quantizer, which therefore adds a random el- sequence also completely determines the input to the
ement to the quantization process. Because it is a special, but SDM.
1214 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 52, NO. 6, JUNE 2005

Second, if a solution to (5) exists, one can define as a matrix is given by . The th element
minimum norm solution to (5) [16] of the last column of is given by

(24)
(16)
Because of the special structure of , we know that
The solution is characterized by the fact that the norm , where is the Kronecker delta. Hence, we have
is the least of all norms of other solutions to .
Though one part of the necessary conditions for a LC has been (25)
solved (5), the set of inequalities represented in (9) still needs to
be solved. For each inequality in (9), one can write an equality which again is a matrix with the last column . By
which represents the conditions under which the constraint is on induction, inherits this special structure, too.
the edge of being violated Therefore, the last column of equals ,
and, hence, the null space of contains at least
. Because for these order SDMs the rank of
(17) equals , the null space is .
This represents an -dimensional hyperplane which bisects Referring to Fig. 2 to see the implication of this, it means
the dimensional space. The point where this surface intersects that the state of the last integrator can be altered over a range
the line defined by (11) represents the boundary where a LC of , without breaking up the LC. However, the effect of
length is on the verge of violating the constraint. This changing the last integrator state is nothing other than adding
point is given by solving for in the equation an offset just before the quantizer. For example, when the last
integrator state is changed by an amount , this is equivalent to
(18)
adding a value to the input of the quantizer with .
Equation (18) defines a distance from the initial point Hence, this approach provides the means to define a minimum
at which the constraint is on the edge of being violated. De- disturbance, just before the quantizer, which is necessary to
pending on the sign of , either (sign posi- break up a LC in an SDM with a dc pole for the last integrator.
tive) or is required in order to fulfill the constraint. 2) Last Integrators With Poles Away From DC: In case the
The set of constraints, (18), can be divided into two categories, last two integrator sections form a resonator, the last column of
and , of feasible , depending on the sign of will be of the form , and, clearly,
the null space does not have the simple shape anymore as in the
previous section. In fact, if the feedback coefficient in the last
two integrator sections equals , it can be shown that to very
if then (19)
good approximation the null-space is given by
if then (20)
(26)
Defining
Hence, in order not to disturb a LC when changing the last inte-
(21) grator section, the third integrator state should also be changed.
Although it is not as easy to determine the exact minimum
(22) disturbance that needs to be applied just before the quantizer
(i.e., change the last integrator state) before the output bit
provides an interval for a feasible
changes sign, it is still possible to define a disturbance which
(23) is equal or larger than this minimum amount. One can do this
under the assumption that , in which case one can apply
Obviously, when , there is no feasible solution, and the theory from the previous section. As typical values for are
the LC cannot exist. of the order of (see Appendix C), validity of the approx-
We will now investigate the nature of the disturbance that imation is asserted. This approximation slightly overestimates
can be applied to the SDM, before the LC breaks up. We will the minimum disturbance required to break up a LC, and thus
separate the two situations for SDMs with dc poles: 1) SDMs represents a pessimistic estimate.
with (more than) one dc pole, where the last integrator section
creates a pole at dc (in Fig. 2 this corresponds with ); B. No DC Poles
and 2) SDMs with (more than) one dc pole where the last two A special situation arises when the SDM has no dc poles. In
integrator sections create a resonator with poles away from dc that case, the null space of is zero. There is only
(in Fig. 2 this corresponds with ). one solution to (5). If this solution also complies with all
1) Last Integrators With DC Poles: The question that needs inequalities (9), it results in a LC. Because the null space is zero,
to be answered, is what the null space in (11) looks like. For the any change of the integrator states would result in a break-up
current case, the last two integrator sections create two dc poles, of the LC. A relevant question that remains, however, is how
which translates to the fact that the last column of the transition long it would take before the bit-pattern is changed from the
REEFMAN et al.: DESCRIPTION OF LCS IN SDMs 1215

LC pattern; in other words, what freedom does one have when causes a bit-flip with respect to the ideal LC pattern, causes
the only requirement is to fulfill (9). This will be the subject break-up of the LC. An approach based on perturbation theory
of Section IV. Note, that the system of inequalities itself would will be followed in order to determine when such a bit-flip will
lead to the same solution as (5) would after an infinite amount occur.
of time (see Appendix A). For a given LC of length , the states at clock cycle are
given by
C. Dither
The basic principle of “dithering” (adding random offsets to (30)
the quantizer) is sketched in Fig. 2. The addition of dither repre- To have some idea about stability of LCs, the original state vari-
sents a special case of LC disturbance, since it does not directly able will be perturbed by an amount
influence the integrator values. The only way in which dither
can break up a LC is by changing the sign of the input to the (31)
quantizer, causing it to create a bit-flip in the LC output. The
minimum amplitude of the dither that is necessary to cer- The growth of a disturbance in the state variables after
tainly break up a LC, is easily determined as periods (and, hence ) of the LC is given
by
(27)
where the dependence of the minimum dither level on the ini- (32)
tial states is explicitly indicated. In a typical situation, where To analyze (32), a Jordan decomposition [16] of is created,
dither according to a certain (e.g., rectangular) pdf spanning a which is defined as1
width is applied, all dither values with amplitude less than
are without any effect. Because of its dependence on (33)
the initial states of the SDM, (27) is not the most convenient
where is a Jordan matrix of the form
expression to determine an appropriate dither level. Preferably,
one would have the expression that provides the maximum of if
over the initial states. In Section III-A1), it was derived if (34)
that for SDMs with the last integrators having their poles at dc, otherwise
the value of the last integrator could vary over a range with the th eigenvalue of the transition matrix . The main
without breaking up the LC. The interpretation is that advantage of this decomposition is that it provides a compact
representation of repeated application of as
if then
(28) (35)
Because the quantizer input bears a linear relation to , it
means that the minimum amplitude dither , needed to break where . From this expression, it is evident that for
up a LC, is maximized over all when , SDMs with eigenvalue magnitudes , multiple applica-
and thus tion of will result in exponential growth of the disturbance.
When , on the other hand, exponential decay will occur.
(29) The effect of a disturbance will be studied in the next sections,
both for SDMs with all poles at dc (all eigenvalues ),
For most SDMs, the value can be easily determined using and for SDMs with resonator sections and eigenvalues .
results obtained previously, without resorting to (27). For the
SDMs which have a resonator section as last integrators, but also A. Only DC Poles
have dc poles, as discussed in Section III-A2), the null-space is In the special case where all the poles of the loop filter are at
not exactly equivalent to a mere change of the last integrator dc, all eigenvalues , as a result of which only polynomial
value. However, for typical SDMs, the value of the feedback growth can occur. In particular, when the eigenvalues are all
coefficient of the last resonator is much less than 1, and there- unity, the result for can be written as
fore the null space is almost equivalent to a change of the last
integrator state. As a result, one can treat such SDMs in exactly if
the same way for determining a lower bound to the minimum if
(36)
amount of dither. otherwise.
For SDMs without dc poles, however, the null-space has di-
mension zero and the methods outlined above cannot be used For an SDM with only dc poles, the transition matrix is ex-
anymore. In this case, the only option is to determine the min- actly in the shape of this Jordan block, with eigenvalues equal
imum amount of dither through (27). to 1, and no further decomposition is necessary (SDMs which
exhibit poles in the loopfilter, are not in Jordan form). In order to
IV. STABILITY ANALYSIS OF LIMIT CYCLES determine when a LC will be broken up, the disturbance
To determine whether a LC is stable, the same assumption 1Alternatively, the Schur decomposition could be used, which, for the types
made in Section III.C will be made that any disturbance that of SDMs under consideration, is less practical.
1216 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 52, NO. 6, JUNE 2005

at the quantizer input that results in a bit-flip must be deter- of the series of resonators will continue to increase steeply,
mined, akin to the discussion in Section III-C. From (2) until the maximum of the sine wave with the lowest frequency
is reached. This maximum occurs when
(37)

Using (36), and making the approxiation that , the (42)

polynomial divergence of a modulator of order can be ap-
proximated by For practical of about , this means that is of the order
of 100 when this maximum is achieved. The exponential factor
has typically grown a negligable 20% by then, which shows
that when has not been reached, LC break up can only
be achieved through the exponential growth. Only for ,
the rate of growth of the exponential divergence equals that of
the polynomial growth.
.. If an SDM consists of a cascade of simple integrators and
. resonators, the output will be a multiplication of the polynomial
(38) divergence as described in Section IV-A, and the oscillatory be-
havior as described above.
where and likewise for . Hence, if the first
integrator is disturbed, the number of LC periods it takes
C. Implications
before a LC is broken up can be approximated by
From (38) and (42), two important observations can be made.
Firstly, because in both cases will be dominated by the
(39)
highest order terms, it will be the weighting coefficient of the
last integrator that determines the rate at which grows.
where is the critical value of where the LC is broken
Hence, more aggressive noise shapers, which have relatively
up. From Section III-A, we have that
large weight at the high order integrators, will exhibit much less
(40) stability with respect to LCs as mild noise shapers will. Sec-
ondly, the growth rate of is in both cases increasing with
where the precise value of depends on the initial state . , which means that long LCs are much less stable, and hence,
much less likely to occur, than short LCs.
B. Resonator Sections Finally, it is worth mentioning that disturbing the SDM just
In case of resonator sections, it proves to be slightly more before the quantizer (dithering the quantizer), is apparently less
difficult to obtain a general algebraic expression for the rate of effective than disturbing the SDM at any other position. Placing
growth of . An expression for the effect of a disturbance can the disturbance anywhere else guarantees that the LC will al-
be obtained from the difference equations describing a resonator ways be broken up, no matter how small the disturbance is.
section, though. From (73) in Appendix B, it is shown that, for
small , the growth of a disturbance in a single resonator section
V. NUMERICAL RESULTS
can be described by
The results of the work detailed in the preceding part, have
been used to obtain some results on several different noise
transfer functions [(NTFs), all Butterworth design], which have
(41) been implemented in feedforward SDMs. The SDMs, all with
where is the input at clock cycle to the resonator, and an oversampling ratio of 64, have been chosen to illustrate
is the output of the resonator (or, effectively, the output the difference in behavior for various SDMs with aggressive
of the last integrator of the resonator) at clock cycle ; noise shaping and mild noise shaping. In the case of most
is the feedback coefficient in the resonator section (see also aggressive noise shaping, the NTF has a high corner frequency
Fig. 2). The time instant is (a bit arbitrarily) taken as of kHz. Although this SDM displays excellent noise
a moment where the bit output pattern coincided with the bit suppression in the baseband, its stability is severely compro-
output pattern corresponding to the LC. It can be observed that mised due to which it is of hardly any practical use. The SDM
the original input to the resonator, , appears at the output, with mild noise shaping has its NTF with corner frequency at
multiplied by , in addition to a sine or cosine, which is of 80 kHz. The naming convention is such, that the first part of
the same order of magnitude as . The output of a series the name of the SDM reflects its NTF corner frequency, and
of resonators, each characterized by a feedback coefficient , the last part is either ”a” (meaning with resonator sections,
will thus consist of a superposition of sinusoidal signals, with i.e., and ) or ”b” (meaning without resonator
frequency , each of them exponentially diverging sections, i.e., ). A more detailed description of all
as . The input to the first resonator will appear at the the SDMs used in this paper, along with full descriptions of
output of the last resonator, multiplied by a factor the SDMs with NTF corner frequencies at 120 and 80 kHz are
if resonator sections are cascaded. From , the output given in Appendix C.
REEFMAN et al.: DESCRIPTION OF LCS IN SDMs 1217

Fig. 3. Occurrence of LCs of SDM 80a and 80b. Fig. 4. Occurrence of LCs of SDMs 80a, 100a, 120a, and 160a.

In the following sections, we will discuss results on static

and dynamic behavior of SDMs. Feedforward topologies are
presented to judge how the implementation topology influences
the LC behavior.

A. Static Behavior
In Fig. 3, the occurrence of LCs for SDMs with and without
resonators is presented as a function of LC length. This occur-
rence has been obtained by generating all independent bit pat-
terns for a given length, and checking whether this pattern could
represent a specific LC with the theory presented in Section III.
Fig. 3 represents all independent LCs that could exist with
lengths ranging from 3 to 30 bits. All possible dc values are
represented by these LCs, and though some of these LCs theo-
retically exist, they cannot occur in practice. For example, it is
possible to define an LC corresponding to an input of 0.9, where Fig. 5. Number of LCs for a fixed LC period of 24, as a function of the NTF
none of the SDMs studied would be capable of representing this corner frequency used in the SDM design.
dc level without running into instability. It is apparent immedi-
ately from Fig. 3, that the presence or absence of resonator co- illustrate this behavior, the dependence of the number of LCs at
efficients for the SDM with NTF corner frequency at 80 kHz is given LC length ( ) on the corner frequency of the Butter-
immaterial to the number of LCs that can occur. The same is ap- worth NTF design is given in Fig. 5. This clearly illustrates the
proximately true when comparing other SDMs with and without increase of the number of LCs with increased aggressiveness of
resonator section. However, a significant difference is displayed the SDM, and also shows that for highly aggressive SDMs the
when comparing the aggressive SDM 120a and the nonaggres- number of possible LCs is virtually constant.
sive SDM 80a in Fig. 4. Rather counterintuitive, the SDM 120a Qualitatively, these observations can be explained on the
displays more LCs than SDM 80a; one would expect the reverse basis of the phase characteristic of the loop filter. To sustain a
to be true, as experimental evidence usually proves stable SDMs LC perfectly, the phase shift for all frequency components of
more susceptible to LCs than aggressive SDMs. The SDMs all the LC needs to be (the feedback loop accounts for another
show an initial steep growth of the number of LCs, followed by factor , summing up to the required corresponding to a
a transition to a region of less steep growth. Based on pure per- delay of one period). For high frequencies, down to the corner
mutations, one would expect the number of LCs to grow propor- frequency of the loopfilter, its phase shift deviates relatively
tionally to . While, indeed, exponential growth of the number little from . Below that frequency, it starts to deviate strongly
of LCs is observed, from a numerical fit the initial growth for from , corresponding to the fact that the likelihood of a long
SDM 120a is proportional to , and for SDM 80a . LC with significant low frequency content being sustained is
Above the cross-over point, the growth is proportional to , low. This causes the rate of growth of the number of LCs to
and this is approximately true up to the largest LC investigated reduce above the loopfilter corner frequency, cf. Fig. 4. Also,
for all SDMs independent of their aggressiveness. Also, the fre- the more aggressive the loopfilter, the less deviation from for
quency of the cross-over point appears to be coincident with the high frequencies. Hence, an aggressive loopfilter will sustain
LC period that corresponds to the corner frequency of the But- more LCs with relatively large high-frequency content than a
terworth high-pass filter that was used in the design. To further nonaggressive filter, cf. Fig. 5.
1218 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 52, NO. 6, JUNE 2005

Fig. 8. Occurrence of LCs as a function of the number of 1’s in the LC, for
Fig. 6. Relative occurrence of LCs of SDM 120a and 80a with respect to all various SDMs. Zero dc level corresponds to 15 1’s, 20 1’s corresponds to a dc
permutations of bits. level of one third, etc..

than those of the less aggressive SMD 80a. Again, this is quite
counter-intuitive as we expect aggressive SDMs to be less sus-
ceptible to LCs. Also, we can see that there is a very stable LC
occurring around LC length 22 for SDM 120a, and for LC length
32 for SDM 80a. Upon investigation of these LCs, it appeared
that they consist of a series of 11 1s followed by 11 s for SDM
120a, and likewise 16 1s and 16 s for SDM 80a. This corre-
sponds to a square wave of frequency 120 kHz and 80 kHz, re-
spectively, which are exactly the corner frequencies of the NTF
design of the SDMs. Although not shown, identical behavior
occurs for other SDMs. In practice, however, these LCs require
huge initial integrator states that could never occur. Long before
such an integrator state could be reached in real operation, the
SDM would have reached a state with unbounded state variables
where the output bit pattern would not reflect the input signal
anymore (see also [1] for a discussion on this phenomenon). As
Fig. 7. Minimum level of dither needed to break up a LC corresponding to a a result, if the SDM has been forced into this LC, the SDM runs
dc input 0 .
unstable upon the slightest disturbance of the integrators.
This is to be contrasted with the LC behavior for other LC
While the absolute number of LCs increases rapidly, relative lengths. The shortest LC, the sequence , appears to be by
to the number of possible permutations of s and s it is far the most stable (disregarding the previously discussed LCs)
reducing rapidly as is demonstrated in Fig. 6. The total number for both SDMs. For longer LCs, the amount of dither needed for
of permutations for an LC of length is approximately break-up decreases to a minimum value close to the peak, after
given by [3] which the LC becomes more stable. All these LCs consist of the
sequence , which repre-
(43) sents the minimally possible deviation for the simple
sequence. While these most stable LCs slightly increase in sta-
The division by corrects for the fact that of all permuta- bility for longer LCs, on average the amount of dither necessary
tions, exactly represent a cyclicly shifted version of the same for break-up decreases. This is indicated in crosses and squares
basic LC. To obtain the exact number of irreducible LCs, cor- for SDM 120a and 120b, respectively, in Fig. 7. The average
rection should be made too for the number of LCs that are a amount of dither is defined as the average of the minimum dither
concatenation of smaller LCs. However, for reasonable , this levels that are needed to break up the individual LCs. Again,
number is much smaller compared to and thus ignored. In we see that SDM 120a presents LCs that are in general more
Fig. 7, the minimum dither level that is needed to certainly break stable than those of SDM 80a. At LC lengths of 42, the average
up the most stable LC is depicted. In plusses (“ ”) and stars amount of dither is reduced to about 0.03 and 0.017 for SDM
(“ ”), the most stable LC for SDM 120a and 80a, respectively, 120a and 80a, respectively, which is consistent with the intu-
for dc input is depicted. While slightly more stable LCs can ition that longer LCs represent more boundary conditions to be
sometimes be found for non-dc inputs, this does not represent fulfilled and are thus more easy to break up. Another interesting
a practical situation. The first interesting observation is that the characteristic to study is the relative preference of the SDM for
LCs for the aggressive SDM 120a are more stable against dither LCs of a certain dc level. These results are displayed in Fig. 8.
REEFMAN et al.: DESCRIPTION OF LCS IN SDMs 1219

Fig. 9. Disturbance of identical LCs for SDM 120b and 80b (with all Fig. 10. Disturbance of LCs of length 10,20 and 30 for SDM 80a due to a
0
eigenvalues equal to 1) due to a small disturbance ( 120 dB) on the first 0
small disturbance ( 120 dB) on the integrator states. Depicted is the quantizer
integrator state. Depicted is the quantizer input until the LC breaks up. input until the LC breaks up. The dotted line represents the exponential growth
that is expected due to the (largest) eigenvalues of the transition matrix.
Confirming the results in Fig. 4, in general SDM 120 exhibits
many more possible LCs for any dc level than SDM 80. Also, dicted at the 38th period for SDM 80b, and at the 29th period for
we see that (as anticipated) the number of LCs is identical for SDM 120b, in good agreement with the observed behavior. The
a certain dc level and the negative dc level. However, whereas same trend is observed in Fig. 10 for the SDMs 120a and 80a,
SDM 120 has strong preference for LCs with small absolute dc which both display eigenvalues ; the largest eigenvalue, cor-
level, SDM 80 apparently has little preference! Moreover, where responding to the pole at 20 kHz, is , which
SDM 80 displays little dependence on the presence of resonator has a norm equal to 1.000 90. The dotted line in Fig. 10 rep-
sections, SDM 120 shows, especially for the smallest dc levels resents the exponential growth that is expected on basis of this
some dependence, displaying most LCs when no resonators are eigenvalue. As a result, displays both oscillatory and expo-
present. The reason for this behavior is unclear. nential growth as can be inferred from Fig. 9. However, a dras-
tically larger number of LC periods passes, before it is broken
B. Dynamic Behavior
up, in line with the expectations mentioned in Section IV. The
In Fig. 9, the effect of a small disturbance of the integrator first maximum in occurs, both for SDM 80a and 120a, at the
states on a LC is illustrated. It depicts the growth of , which 15th period, exactly equal to the prediction of (42). The pre-
is the deviation of the quantizer input from its ideal input, as de- dicted amplitude at these maxima equals 0.000 25 and 0.0017
fined in (37). The LC studied was the most stable of length 8, for SDM 80a and 120a, respectively, which is about a factor of
i.e., followed by a sequence of 2 pairs. 1.5 too low with respect to the simulation result. This difference
A disturbance of dB ( ) was applied to the first inte- is easily explained because of the fact that the input to the cas-
grator at time instant in order to break up the LC. The cade of resonators is a constant (due to the action of the first inte-
effect of such a disturbance on the output signal during normal grator). This means that the last term in (41) is nonzero, which
SDM operation is very small; in fact, it is much less than the effect is ignored in determining the maximum value. Because
effect that sufficiently dithering the quantizer would have had. the first maximum is still far from , the total duration of
Studying the case where all transition matrix eigenvalues equal the LC is determined by the exponential growth. The predicted
1, we can clearly see that for SDM 120b, which is the more ag- duration of the LC is 807 periods for SDM 80a, and 714 periods
gressive one, the deviation from the ideal quantizer input in- for SDM 120a, which is both about 5% shorter than observed in
creases steeply. This, in turn, results in early break-up of the LC. simulation.
For the nonagressive SDM 80b, this increase is much less steep.
As both SDMs have all loop filter poles at the unit circle, the rate
VI. CONCLUSION
of growth of is polynomial. Even though the maximum devi-
ation , for which LC break-up occurs for this SDM, is much This work is an attempt to construct a general theory de-
less than for SDM 120b, it takes a 30% longer time for it to break scribing LCs in 1-bit SDMs, and to provide the designer with
up compared to the same LC in SDM 80b. Note, that the fact that tools other than numerous simulations to obtain an insight into
the maximum deviation for which break-up occurs is larger for typical LC behavior of SDMs. It has been proven that, under al-
the least aggressive SDM, is in correspondance with the results most all circumstances, LC behavior is observed in the output
on dither as presented in Fig. 7. Thus, aggressive noise shapers if and only if a LC occurs in state space. It has been shown that
need more dither in order to break up a LC when this dither is LC behavior can occur in a wide variety of situations.
added to the quantizer, and need less dither when the dither is In Section III.A, a recipe was given whereby, for constant
added to the input of the SDM to break up a LC. Upon substitu- input, all LCs of a given period can be found for any SDM with
tion of the relevant parameters in (39), break-up of the LC is pre- at least one pole at dc. Equation (16) provides a least squares
1220 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 52, NO. 6, JUNE 2005

solution to the LC conditions. If the constraint equations, (9) structures, and feedback SDMs have also been investigated by
and (15), are also satisfied, then this is an exact solution. Equa- the authors [7]. The use of constant input is a highly relevant
tion (18) may then be solved to find the exact set of initial con- situation, since these LCs are known to be problematic and
ditions, (11), which give rise to this LC. This same procedure easily observed. However, SDMs are intended to be used pri-
can be applied when the SDM has no dc poles. However, it be- marily with input signals of a limited bandwidth. The dynamic
comes simpler in this situation since (5) can now be solved di- behavior of SDMs under periodic or noisy input is still an
rectly, and (9) is the only constraint. The essential difference unknown area, and is currently a research topic pursued by the
between these two situations is that, if constraint equations are authors.
satisfied, SDMs with dc poles will exhibit a line of initial condi-
tions which give rise to a LC, whereas SDMs without dc poles APPENDIX A
will exhibit a unique solution. One immediate consequence of PROOF THAT EQUALITIES ARE A SUFFICIENT
the initial condition dependence is that, for an SDM of order CONDITION FOR A LIMIT CYCLE
with dc poles, states need to have a well-defined value, and
In this appendix, we will prove that the set of inequalities (9)
all states need to have a well-defined value for SDMs without
leads to the same solution as the set of equalities (5) would, i.e.,
dc poles. This makes LCs for higher order SDMs (which typi-
whether the LC condition observed at the output of the SDM
cally also exhibit more aggressive noise shaping) less likely to
occur, especially when they do not exhibit poles at dc. (44)
It is postulated that the most stable LC is that which is a se-
quence of pairs, followed by a single leads to the existence of a periodic orbit in state space as known
quartet. It has also been shown, that the classical approach of in stability analysis
dithering the quantizer may not be the optimal way of removing
LCs. Adding a small disturbance to an integrator state is far (45)
more efficient and will always result in break up of the LC. The
noise penalty is rather limited, as the input disturbance can be For example, if the output sequence happens to be a periodic
made as small as or dB. Furthermore, it has been sequence of period , one could ask the question whether there
shown that very small changes in an SDMs structure can have exists a possibility that an initial state does not return to this
significant effects on the rate of growth of any disturbance to a value after propagation over cycles, but to a different state
LC. SDMs with only dc poles will exhibit polynomial growth, vector . If this state vector generates the same output sequence
whereas the inclusion of resonator sections or other modifica- again, etc., we have a LC without fulfilment of (5). To that end,
tions to the structure may yield exponential growth. However, if we look at the propagation of the state variables (3), after a large
these modifications result in the transition matrix having com- number of cycles. We will further assume that is
plex conjugate pair eigenvalues, then the exponential growth is invertible; when it is not we will define a new transition matrix
exhibited as the disturbance spiraling away from initial con- as
ditions. Thus, this exponential growth may take significantly
(46)
longer to break up the LC than the polynomial growth which
occurs without resonators. Therefore, in general, SDMs without where is the original transition matrix, and is the unit ma-
resonators are less susceptible to LCs. trix. Now is invertible by definition; at the end of the
An important characterization given the goals of SDM de- analysis we will than have to take the limit to obtain the
sign, is distinguishing LC behavior for SDMs with different final result.
noise shaping characteristics. Intriguingly, SDMs with aggres- From (3), we subsequently determine the states at the
sive noise shaping can sustain many more different LCs than th clock cycle as
SDMs (of equivalent order) with mild noise shaping, and are
also more robust against dithering the quantizer. Also, it has
been shown that LCs of a long period, even though the number (47)
of LCs grows exponentially, are much more sensitive to a small
disturbance than a short LC. Likewise, it can be proven that The summation can be written as two nested summations
SDMs with aggressive noise shaping are more sensitive to small
disturbances than mildly noise shaping SDMs—even though the
latter exhibit a much smaller number of sustainable LCs. This
is corroborated by the general experimental observation that ag- (48)
gressive noise shapers are less susceptible to LCs than mild
noise shapers. As a result, dithering the quantizer as a means to
Because is a LC and is constant, we have, by def-
remove LCs should be discouraged. Adding (even tiny) amounts
inition, , and thus we can write
of, perhaps shaped, noise to the input of the SDM is far more ef-
fective in achieving the same goal.
It should be noted that all the results and observations pre-
sented here apply only to feedforward SDMs with constant (49)
input. However, the technique may be generalized to other
REEFMAN et al.: DESCRIPTION OF LCS IN SDMs 1221

The terms not dependent on can be moved outside the

summation

(50)
Fig. 11. Single resonator section, as used in an SDM.

and the summation over can be written more simply as is independent of the invertibility of . The definition of
in (55) also turns out to be a familiar result, as with
it is identical to (6) in Section II.B.

(51) B. Case 2: Corresponds to Eigenvalues With Norm

Because the sum over represents a constant, we define Smaller Than 1
If, on the other hand, contains only directions which
correspond to eigenvalues of , these directions will be
(52) reduced to zero if . We thus obtain as a result

in line with earlier definitions in Section II.B. With this defini- (59)
tion, we can write (49) concisely as
which is identical to (57) when . However, we do
not have the result that . Thus, in this case, we have
(53)
a possibility that an initial state does not return to this value
after propagation over cycles, but to a different state vector
Realizing that the finite sum represents a geometric series, and , and that this state vector generates the same output sequence
because we have defined such, that is by definition again. After a long number of LC periods, however, con-
invertible, (53) can be expressed as follows: verges to a unique value, such that
which is a situation identical to Case 1.
(54)

For convenience, we will further define APPENDIX B

SOLUTION TO THE DIFFERENCE EQUATIONS OF
A RESONATOR SECTION
(55)
In order to be able to determine the effects of a resonator sec-
due to which we can further compress (53) as tion in an SDM, we will solve the difference equation describing
such a system. Whereas a linear algebraic approach, such as
(56) used in the main text of the paper, will give the same results,
it proves to be a rather rather cumbersome excercise, for which
In order for a LC to be stable, must be reason we refrain from presenting this approach. The resonator
bounded. We will discriminate two cases. system that we study, is depicted in Fig. 11. The input to the
system is , and the system output that we will study, is labeled
A. Case 1: Corresponds to Eigenvalues With Norm in Fig. 11. The output labeled can be obtained in an almost
Larger Than 1 identical way as described below.
If contains directions which correspond to eigen- The difference equations describing the resonator output
values with a norm larger than 1 of , is are given by
bounded only when
(60)
(57)

leading to different realizations of a resonator section (for example, with

nondelayed integrators etc.) will have a slightly different but
(58) comparable difference equation. We will seek the solution to
(60) as
which is the conjecture on which the results in Section II.B are
based. (61)
Upon applying the definition of (46) to (56), we see that we
can safely take by setting , as a result of which (57) where is the particular solution.
1222 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 52, NO. 6, JUNE 2005

We will now first want to find the homogeneous solution It is the boundary conditions that will determine the constants
, that is, the solution to the substitution of and .
in (60) The first obvious requirement is that be real, that is,
. Further we have that either
(62)
(71)
which is identical to
when a step function is applied to the input, or
(63)

Solving for , we obtain two independent solutions and (72)

to
when a gradually increasing function is applied to the input. It
is easily verified that boundary condition (71) results in

(64) (73)
while the second set of boundary conditions (72) results in
This homogeneous solution can also be obtained rather easily
from the algebraic approach, as the eigenvalues from the transi-
tion matrix describing the resonator are given by
. (74)
To obtain the particular solution , we will substitute
in (60). This results in The first solution (73) is more realistic in the sense that it is very
unlikely that , and is the solution that leads to (41) in
(65) the main text upon the assumption that varies only slowly,
leading to the substitution .
For slowly varying , this can be approximated by

(66) APPENDIX C
DESCRIPTION OF SIGMA DELTA MODULATORS
where is the second derivative of . A solution for The feedforward SDMs used in Section V are all fifth order,
can now be found when both and are expanded in a and are referred to by a code which gives their corner frequency
Fourier series and a letter “a” if they include resonators, and “b” if not. They
are all of the type displayed in Fig. 2

(67)

kHz
When these Fourier expansions are substituted in (66), we ob-
kHz
tain the Fourier coefficients and
kHz
kHz
(68)
kHz
kHz
For example, if is a constant, or a slowly varying function,
a good approximation for is given by

The design of the SDMs is such that the NTFs of the

(69)
SDMs are all of the Butterworth high-pass type if . For
SDM120a, the NTF has a dB point at 120 kHz; for SDM80a,
The general solutions to (60) now read this point is at 80 kHz. Thus, SDM80a represents a much less
aggressive noise shaper than SDM120a. The coefficients for
these two SDMs are tabulated in the table shown at the bottom
(70) of the page.
REEFMAN et al.: DESCRIPTION OF LCS IN SDMs 1223

REFERENCES Josh Reiss has Bachelor’s degrees in both physics

and mathematics, and the Ph.D. degree in physics
[1] S. Norsworthy, R. Schreier, and G. Temes, Delta-Sigma Converters,
from the Georgia Institute of Technology, Atlanta.
Theory, Design and Simulation. New York: IEEE Press, 1997. In June 2000, he accepted a research position in
[2] F. de Jager, “Delta modulation – A method of PCM transmission using
the Audio Signal Processing Research Laboratory,
the one unit code,” Philips Res. Rep., vol. 7, pp. 442–466, 1952. King’s College, London, U.K. He has since become
[3] L. Risbo, “Sigma–delta modulators: Stability analysis and optimiza- a Lecturer with the Centre for Digital Music in the
tion,” Tech. Univ. of Denmark, Denmark, 1994.
Electronic Engineering department at Queen Mary
[4] S. Hein and A. Zakhor, Sigma–Delta Modulators: Nonlinear Decoding University of London, London, U.K. His research
Algorithms and Stability Analysis. Norwell, MA: Kluwer, 1993.
interests include nonlinear dynamics, audio and
[5] S. Ardalan and J. Paulos, “An analysis of nonlinear behavior in delta- music processing, and music retrieval systems. He
sigma modulators,” IEEE Trans. Circuits Syst., vol. CAS-34, no. 5, pp.
is the vice-chair of the Audio Engineering Society Technical Committee
593–603, May 1987. on high-resolution audio, and the Program Chair of the 6th International
[6] A. J. Magrath, “Algorithms and architectures for high resolution
Conference on Music Information Retrieval (ISMIR’05).
sigma–delta converters,” Univ. of London, London, U.K., 1996.
[7] D. Reefman, J. Reiss, E. Janssen, and M. Sandler, “Description of limit
cycles in feedback Sigma–Delta modulators,” in Proc. AES 117th Con-
vention, 2004, San Francisco, CA, Oct. 28–31, 2004.
[8] S. Mann and D. Taylor, “Limit cycle behavior in the double-loop band-
pass Sigma–Delta A/D converter,” IEEE Trans. Circuits Syst. II, Analog
Digit. Signal Process., vol. 46, no. 10, pp. 1086–1089, Oct. 1999.
[9] V. Friedman, “The structure of the limit cycles in Sigma–Delta modula-
tion,” IEEE Trans. Commun., vol. 36, pp. 972–979, 1988.
[10] N. Bridgett and C. Lewis, “Effect of initial conditions on limit cycle per-
formance of second order sampled data Sigma–Delta modulator,” Elec-
tron. Lett., vol. 26, pp. 817–819, 1990.
[11] D. Reefman and P. Nuijten, “Editing and switching in 1-bit audio
streams,” in Proc. AES 110th Convention, Amsterdam, The Nether- Erwin Janssen was born in The Netherlands in
lands, May 12–15, 2001. 1976. He received the M.Sc. degree (cum laude) in
[12] D. Hyun and G. Fischer, “Limit cycles and pattern noise in single-stage electrical engineering, with an additional degree in
single-bit delta-sigma modulators,” IEEE Trans. Circuits Syst. I, computer science, from the University of Twente,
Fundam. Theory Appl., vol. 49, no. 6, pp. 646–656, Jun. 2002. Enschede, The Netherlands, in 2001.
[13] A. Oppenheim and A. Shafer, Discrete-Time Signal Processing. En- He joined Philips Research in 2001, to work as a
glewood Cliffs, NJ: Prentice-Hall, 1989. Research Scientist in the Mixed-Signal Circuits and
[14] D. Reefman and E. Janssen, White Paper on Signal Processing for Systems Group. Since then, he has been working on
SACD. Eindhoven, The Netherlands: Philips, 2003. various aspects of signal processing for super audio
[15] P. Steiner and W. Yang, “A framework for analysis of high-order sigma- compact disk. His research interests include audio
delta modulators,” IEEE Trans. Circuits Syst. II, Analog Digit. Signal signal processing and sigma–delta modulation.
Process., vol. 44, pp. 1–10, 1997.
[16] G. Golub and C. Van Loan, Matrix Computations, 3rd ed. Baltimore,
MD: The Johns Hopkins Univ. Press, 1996.

Derk Reefman received the Master’s degree in

chemistry and the Ph.D. degree in physics from the
university of Leiden, Leiden, The Netherlands, in
1989 and 1993, respectively.
He joined Philips Research, Eindhoven, The
Netherlands, in 1993, and has been active in various
aspects of X-ray diffraction. Since 1998, he com-
bined hobby and work and joined the Mixed-Signal Mark Sandler was born in 1955. He received the
Department in Philips Research where he works on B.Sc. and Ph.D. degrees from the University of
digital and analog aspects of high-quality audio, in Essex, Essex, U.K., in 1978 and 1984, respectively.
particular super audio compact disk and Class-D He is a Professor of Signal Processing and Di-
amplifiers. Late 2003, he switched to the field of power management and rector of the Centre for Digital Music at Queen
power conversion, where his current areas of interest include integrated dc–dc Mary University of London, London, U.K., where
conversion and high bandwidth dc–dc conversion. He has significant experience he moved in 2001 after 19 years at King’s College
in system design and nonlinear dynamic system theory and he owns several London, U.K. He was Founder and CEO of Insonify,
patents in these fields. Ltd., London, U.K., an Internet audio streaming
Dr. Reefman received the “Unilever Research Prijs” for the Best Master’s startup for 18 months. He has published over 250
thesis in chemistry in 1990, and was awarded a prize for the Best Leiden Ph.D. papers in journals and conferences. He is a Fellow
thesis in the field of physics, chemistry, and biology in 1993. He is also the Chair of the Institute of Electrical Engineers (IEE) and a Fellow of the Audio
of the Audio Engineering Society Technical Committee on Audio Recording Engineering Society. He is a two-times recipient of the IEE A. H. Reeves
and Storage Systems. Premium Prize.