Nonlinearity robust successive requantizer

Abstract

A successive requantizer, which serves as a replacement for a ΔΣ modulator in a fractional-N PLL or a DAC, and avoids spurious tone problems, thereby circumventing the tradeoffs that result from reliance on the common approach of making highly linear analog circuitry to avoid spurious tones. A successive requantizer fractional-N PLL of the invention has the potential to reduce power consumption and the cost of commercial communication devices. The successive requantizer performs digital quantization one bit at a time in such a way that the quantization noise can he engineered to have desirable properties such as non-linearity robustness. The successive requantizer is applicable to most high-performance digital communication systems, such as cellular telephone handsets and wireless local and metropolitan area network transceivers.

Description

FIELD

A field of the invention is wireless communications. Another field of the invention is digital to analog converters. Example applications of the invention include high-performance digital communication systems such as cellular telephone handsets and wireless local and metropolitan area network transceivers.

BACKGROUND

Virtually all wireless communication systems require local oscillators for up-conversion and down-conversion of their transmitted and received signals. Most commonly, the local oscillators are implemented as fractional-N PLLs (phase locked loops). Fractional-N PLLs can generate a high quality, high frequency signal from a digital input stream and a lower frequency reference waveform.

The spectral purity of the local oscillator in a transceiver is a critical factor in overall transceiver performance. Communication standards therefore explicitly or implicitly stipulate stringent phase noise requirements on the local oscillators. Most standards dictate the maximum acceptable phase noise power in various frequency bands. Additionally, most standards require that spurious tones in the local oscillator's phase noise be highly attenuated, particularly in critical frequency bands. DACs also have stringent spurious tone requirements in wireless applications, digital audio applications and instrumentation applications. Digital audio and instrumentation spurious tone requirements are particularly critical in many consumer products.

Modern digital-to-analog converters (DACs) and fractional-N phase-locked loops (PLLs) rely upon a digital delta-sigma modulator (ΔΣ modulator) to coarsely quantize a constant or slowly varying digital sequence. The quantized sequence can be viewed as the sum of the original sequence plus spectrally shaped quantization noise that has most of its power outside of a given low-frequency signal band. Ultimately, the quantized sequence is converted to an analog signal and further processed by analog circuitry including a low-pass filter to suppress quantization noise outside of the signal band.

FIG. 1 is a block diagram of a conventional ΔΣ fractional-N PLL. The purpose of the system is to generate an output signal of frequency (N+α)f_ref where N is a positive integer, a is a constant fractional value between 0 and 1, and f_ref is the frequency of a reference oscillator. The system consists of a phase-frequency detector (PFD) and a charge pump 100, a loop filter 102, a voltage controlled oscillator (VCO) 104, a multi-modulus divider 106, and a digital ΔΣ modulator 108. The divider output, v_dlv(t), is a two-level signal in which the nth and (n+1)th rising edges are separated by N+y[n] periods of the VCO output, for n=1, 2, 3, . . . , where y[n] is a sequence of integers generated by the ΔΣ modulator 108. As indicated in FIG. 1B for the case where the PLL is locked, if the nth rising edge of the reference signal, v_ref(t), occurs before that of v_dlv(t), the charge pump 100 generates a positive current pulse of magnitude l_CP with a duration equal to the time difference between the two edges. This increases the VCO control voltage; v_ctrl(t), thereby increasing the VCO output frequency. Alternatively, if the nth rising edge of v_ref(t), occurs after that of v_dlv,(t), the situation is similar except the polarity of the current pulse is negative, which decreases the VCO frequency. This causes the output frequency to settle to f_ref times the sum of N and the average of y[n].

If y[n] could be set directly to the desired fractional value, a, directly then the output frequency of the PLL would settle to (N+α)f_ref, thereby achieving the goal of the fractional-N PLL. Unfortunately, y[n] is restricted to integer values because the divider 106 is only able to count integer VCO cycles whereas a is a fractional value. To circumvent this problem y[n] is designed to be a sequence of integers that average to α. The input to the ΔΣ modulator 108 is α plus pseudo-random least significant bit (LSB) dither, so its output has the form y[n]=α+s[n], where s[n] is a zero-mean sequence consisting of spectrally shaped ΔΣ quantization noise and LSB dither. The dither prevents s[n] from containing spurious tones that would otherwise show up as spurious tones in the PLL's output. Hence, the output frequency settles to an average (N+α)f_ref, as desired, although s[n] introduces phase noise.

For most wireless applications spurious tones can be sufficiently suppressed only with design tradeoffs that significantly degrade other aspects of performance. This is particularly problematic in fractional-N PLLs wherein the input to the modulator usually is a constant and the output sequence from the modulator is converted to analog form and subjected to various nonlinear operations because of nonideal circuit behavior. The common approach to address spurious tones is to make the analog circuitry very linear so that the spurious tones have sufficiently low power for the given application. This limits design options and results in higher analog circuit power consumption than would be required if fewer linear analog circuits could be tolerated.

The tradeoffs tend to increase power consumption and circuit area, limit the choice of reference frequencies, and dictate low PLL bandwidths which preclude on-chip loop filters. They also become less effective in system-on-chip designs as CMOS circuit technology is scaled into the sub-100 nanometer regime. Therefore, the spurious tone problem negatively affects power consumption, cost, and manufacturability of wireless transceivers, and the problem gets worse as CMOS circuit technology scales with Moore's Law.

To address the quantization noise, researchers have described methods of implementing noise-shaped quantizers. See, e.g., A. J. Magrath and M. B. Sandler, “Efficient Dithering of Sigma-delta Modulators with Adaptive bit Flipping,” Electron. Lett, vol. 31, no. 11, pp. 846-847, May 1995; S. H. Yu, “Noise-Shaping Coding through Bounding the Frequency Weighted Reconstruction Error,” IEEE Trans. Circuits Syst. II: Expr. Briefs, vol. 53, no. 1, pp. 67-71, January 2006; D. E. Quevedo and G. C. Goodwin, “Multistep Optimal Analog-to-Digital Conversion,” IEEE Trans. Circuits and Systems I: Regular Papers, vol. 52,no. 3, pp. 503-515, March 2005; I. Daubechies and R. DeVore, “Approximating a Bandlimited Function Using Very Coarsely Quantized Data: A Family of Stable Sigma Delta Modulators of Arbitrary Order,” Ann. Math., vol. 158, no. 2, pp. 679-710, 2003.

Generally, known phase noise cancelling fractional-N PLLs are effective to cancel phase noises that result from mismatches between positive and negative current sources in the charge pump. However, these methods specifically focus on stabilizing noise-shaped coders and do not address the effect of nonlinearities on the quantization error.

SUMMARY OF THE INVENTION

An embodiment of the invention is a successive requantizer, which serves as a replacement for a ΔΣ modulator in a fractional-N PLL or DAC, and avoids the above-mentioned spurious tone problem, thereby circumventing the tradeoffs that result from reliance on common approach of making highly linear analog circuitry to avoid spurious tones. A successive requantizer fractional-N PLL of the invention has the potential to reduce power consumption and the cost of commercial communication devices. A successive requantizer of the invention performs digital quantization one bit at a time in such a way that the quantization noise can be engineered to have desirable properties such as non-linearity robustness. The invention is applicable to most high-performance digital communication systems, such as cellular telephone handsets and wireless local and metropolitan area network transceivers.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1A (prior art) is a block diagram of a fractional-N PLL; FIG. 1B illustrates waveforms in the PLL of FIG. 1A;

FIG. 2A is a block diagram of a preferred embodiment successive requantizer for a fractional-N PLL or DAC;

FIG. 2B illustrates the general form of a preferred embodiment quantization block in the successive requantizer of FIG. 2A in which all variables are integer-valued two's complement numbers;

FIGS. 2C-2E illustrates detail on the sequence generation of the successive requantizer of FIGS. 2A-2B;

FIG. 3A is a block diagram that shows the register transfer level details of the dth quantization block of FIG. 2A, and FIG. 3B illustrates the combinatorial logic truth table of an example implementation of sequence generation logic; and

FIG. 4 is a block diagram of a preferred embodiment prototype fractional-N PLL or DAC.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

It is critical spurious tones in the noise introduced by DACs and fractional-N PLLs have very low power. Unfortunately, spurious tones are inevitable in the phase noise of fractional-N PLLs. In principle, dither applied to a modulator can prevent the quantization noise from containing any spurious tones whatsoever. In practice, digital delta sigma modulators are major sources of spurious tones in oversampling DACs and fractional-N PLLs and are believed by the present inventors to give rise to spurious tones whenever their quantization noise is subjected to nonlinear distortion.

The present invention identifies and addresses a surprising fundamental source of spurious tones in fractional-N PLLs, namely, the digital ΔΣ modulator. The fractional-N PLL is identified as the fundamental source of spurious tones in the PLL's phase noise. This is true even when dither is used to prevent spurious tones in the ΔΣ modulator's quantization noise. Regardless of how dither is applied, spurious tones are induced when the ΔΣ modulator's quantization noise is subjected to non-linear distortion. In a typical ΔΣ modulator used in a fractional-N PLL, the interaction of the constant input and the modulator's first accumulator gives rise to hidden periodicities. Dither provides sufficient randomness to avoid spurious tones in y[n], but not to avoid spurious tones when s[n] is subjected to nonlinear distortion.

This is particularly problematic in fractional-N PLLs wherein the output sequence from the ΔΣ modulator is converted to analog form and subjected to various non-linear operations because of non-ideal circuit behavior. In embodiments of the invention, a successive requantizer replaces the ΔΣ modulator and avoids the spurious tone problems. The successive requantizer can also serve as a replacement for a ΔΣ modulator in digital analog converter, and example embodiment DACs take the same form as a fractional-N PLL.

An embodiment of the invention is a successive requantizer, which serves as a replacement for a ΔΣ modulator in a fractional-N PLL, and avoids the above-mentioned spurious tone problem, thereby circumventing the tradeoffs that result from reliance on common approach of making highly linear analog circuitry to avoid spurious tones. A successive requantizer fractional-N PLL or DAC of the invention has the potential to reduce power consumption and the cost of commercial communication devices. A successive requantizer of the invention to performs digital quantization one bit at a time in such a way that the quantization noise can be engineered to have desirable properties such as non-linearity robustness. The invention is applicable to most high-performance digital communication systems, such as cellular telephone handsets and wireless local and metropolitan area network transceivers.

Preferred embodiments of the invention will now be discussed with respect to the drawings. The drawings may include schematic representations, which will be understood by artisans in view of the general knowledge in the art and the description that follows. Features may be exaggerated in the drawings for emphasis, and features may not be to scale.

Fractional-N PLLs used in communications applications and DACs using ΔΣ modulators ultimately generate analog waveforms. Each such waveform contains components corresponding to digitally generated quantization noise, s[n], and, in the case of fractional-N PLLs, its running sum,

$\begin{array}{cc}t\left[n\right]=\sum _{k=0}^{n}s\left[k\right].& \left(1\right)\end{array}$

Inevitable non-ideal analog circuit behavior causes non-linear distortion. The distortion can be any non-linear function, but for almost all practical applications can be represented by a memory-less, truncated power series. This gives rise to components in the output waveform corresponding to s^p[n] for p=1, 2, 3, . . . , h_s, and t^p[n] for p=1, 2, 3, . . . , h_t, where h_s, and h_t are the highest significant orders of distortion for the given application applied on s[n] and t[n] respectively.

Most communication system standards specify required performance in terms of quantities that can be measured using spectrum analyzers, so the properties of the waveforms typically are quantified in the laboratory using spectrum analyzers. Although the waveforms themselves are considered to be random processes in most cases, spectrum analyzers can only average over time, not over ensemble. Therefore, in such applications the properties of the periodograms of s^p[n] and t^p[n] given by

$\begin{array}{cc}{I}_{s^p,L}\left(\omega \right)=\frac{1}{L}{\sum _{n=0}^{L-1}{s}^p\left[n\right]e^{-j\omega n}}^2\mathrm{and}& \left(2\right)\\ I_{t^p,L}\left(\omega \right)=\frac{1}{L}{\sum _{n=0}^{L-1}{t}^p\left[n\right]e^{-j\omega n}}^2& \left(3\right)\end{array}$

are of particular interest, rather than traditional power spectral density (PSD) functions. It is well known that in certain cases the expected values of the periodograms converge to the true PSD functions in the limit as L→∞, but in the applications mentioned above this is not a requirement, or even relevant to the measured performance. The successive requantization of the invention can be understood with respect to the properties of the periodograms given by (2) and (3).

Preferred Embodiment Successive Requantizer and PLL

FIG. 2A illustrates a preferred embodiment successive requantizer that can replace a ΔΣ modulator, such as the ΔΣ modulator 108 in the FIG. 1A fractional-N PLL. The fractional-N PLL of FIG. 1A as modified by the successive requantizer of FIG. 2A is a type of DAC, as it converts a digital sequence to an analog output. The successive requantizer of FIG. 2A readily serves as the replacement for any converter that uses a delta sigma modulator. The bit sequences illustrated are examples only, and the number of quantization blocks 202₁-202_d will depend upon the number of bits in the input bit sequences and the desired number of bits in the output sequence. The successive requantizer 200 illustrated in FIG. 2A quantizes a 19-bit input sequence received at its input 204 by 16 bits (using 16 quantization blocks 202) to generate a 3-bit output sequence at its output 206.

The input 204 and output 206 of the successive requantizer 200 are integer-valued. For the fractional-N PLL application, the goal is to quantize α, which is a fractional value between 0 and 1, and in this design a is taken to be a constant multiple of 2⁻¹⁶. Therefore, a is scaled by 2¹⁶ prior to the successive requantizer to convert it into an integer. As explained below, the 3-bit integer-valued output of the successive requantizer is y[n]=α+s[n], where s[n] is quantization noise. While the example assumes that each quantization block quantizes its input sequence by one bit, the quantization blocks can quantize their inputs by more than one bit, as will be appreciated by artisans.

As shown in FIG. 2B each quantization stage/block 202_d simultaneously halves its input and quantizes the result by one bit every sample period. The output of the dth quantization block 202_d is x_d+1[n]=(x_d[n]+s_d[n])/2, where s_d[n] a sequence generated within the quantization block. At each time n, s_d[n] is chosen such that x_d[n]+s_d[n] does not exceed the range of a (20−d)-bit (calculated by 4+N−d, where N is the number of quantization blocks) two's complement integer, and the parity of s_d[n] is the same as that of x_d[n]. The range can be generally stated as (K−d)-bit two's complement integer, where the parity of s_d[n] is the same as that of x_d[n], where K is determined by the number of quantization stages, the number of bits quantized per stage, and the logic which generates s_d[n].

The parity restriction ensures that x_d[n]+s_d[n] is an even number so its LSB is zero. Discarding the LSB simultaneously halves the quantization block's input value and quantizes the result by one bit. The resulting quantization noise is s_d[n]/2, so the successive requantizer's overall quantization noise is

$\begin{array}{cc}s\left[n\right]=\sum _{d=1}^{16}{2}^{d-17}{s}_d\left[n\right],& \left(4\right)\left(i\right)\end{array}$ which can be generalized to

$\begin{array}{cc}s\left[n\right]=\sum _{d=0}^{K-1}{2}^d{s}_d\left[n\right].& \left(4\right)\left(\mathrm{ii}\right)\end{array}$

Therefore, s[n] is a linear combination of the s_d[n] sequences, so it inherits the properties of the s_d[n] sequences.

The example presented in (4)(i) is for a specific number of input bits and quantization but can be generalized as in (4)(ii) for different numbers of input bits and different numbers of bit quantizations. Generalized for an input that is a sequence of B-bit numbers, x₀[n], and an output that is a sequence of B-K-bit numbers, x_K[n], where n=0, 1, 2, . . . , is the time index of the sequences. The generalization is shown in FIGS. 2C-2E. The successive requantizer then consists of K quantization blocks 202_K, each of which quantizes its input by one bit, so the successive requantizer quantizes K bits overall. Each quantization block generates a quantization sequence, s_d[n], with the property that x_d[n]+s_d[n] is an even number for each n, where x_d[n] is the quantization block's input sequence. The quantization block adds s_d[n] to x_d[n] and discards the least significant bit (LSB) to implement the 1-bit quantization. Without loss of generality, numbers within the successive requantizer are taken to be integers with a two's-compliment binary number representation. Since x_d[n]+s_d[n] is an even number for each n, its LSB is zero, so discarding the LSB does not incur a truncation error. Hence, the quantization noise of the successive requantizer is a weighted sum of the s_d[n] sequences.

The properties of the quantization noise s[n] can be engineered by appropriate design of the s_d[n] sequences. So far, the only restriction on the s_d[n] sequences is that x_d[n]+s_d[n] must be an even integer for each n and d. In the example successive requantizer 200 this restriction on the s_d[n] sequences is that they must be chosen such that x_d[n]+s_d[n] is a (20−d)-bit two's complement even integer for each n and d. This leaves considerable flexibility in the design of the s_d[n] sequences, which can be exploited to achieve the desired quantization noise properties.

The successive requantizer partially exploits this flexibility to ensure that the running sum of each s_d[n] sequence, i.e.,

$\begin{array}{cc}{t}_d\left[n\right]=\sum _{k=0}^n{s}_d\left[k\right],& \left(5\right)\end{array}$

is bounded for all n, and each s_d[n] has a smooth PSD (power spectral density) that increases monotonically with frequency. This implies that s[n] is highpass shaped quantization noise that is free of spurious tones and the PSD of s[n] is zero at ω=0.

This still leaves flexibility in the design of the s_d[n] sequences which can exploited as described below to ensure that the sequences

(s[n])^p for p=1, 2, 3, 4, 5, and (t[n])^p for p=1, 2, 3,

are free of spurious tones, where t[n] is the running sum of s[n] given by:

$t\left[n\right]=\sum _{k=0}^{n}s\left[k\right]$

The objective is to ensure that the successive requantizer's quantization noise does not introduce significant spurious tones when subjected to the degree of nonlinear distortion expected from the analog circuits within the PLL. Circuit simulations can be used during PLL design to verify that preventing spurious tones from occurring in the (s[n])^p and (t[n])^p sequences is sufficient to achieve this objective.

FIG. 3A is a block diagram that shows the register transfer level details of the dth quantization block, and FIG. 3B illustrates the combinatorial logic truth table. There are many choices of sequences which can yield non-linear robust properties. The truth table of FIG. 3B illustrates one particular example. Each value of s_d[n] is calculated via the combinatorial logic 300 as a function the previous value of t_d[n] provided by D flip flop's 302, the parity of the current value of x_d[n], and the current value of a 4 bit pseudo-random sequence (provided by pseudo random number generator 304), r_d[n], where {r_d[n], d=1, 2, . . . , 16, n=0, 1, 2, . . . } well-approximates independent identically distributed random variables. For this design the range of values taken on by s_d[n] and t_d[n] ares_d[n]ε{−3, −2, −1, 0, 1, 2, 3}, and t_d[n]ε{−2, −1, 0, 1, 2},  (6)

It can be verified that t_d[n] is a discrete-valued Markov random sequence conditioned on the parity of x_d[n]. Whenever x_d[n] is odd the one-step state transition matrix for t_d[n] is given byA_o=[P{t_d[n]=T_j|t_d[n−1]=T_i,o_d[n]=1}]_5×5  (7)

and whenever x_d[n] is even the one-step state transition matrix for t_d[n] is given byA_e=[P{t_d[n]=T_j|[n−1]=T_i,o_d[n]=0}]_5×5  (8)

where P{X|Y} denotes the conditional probability of event X given event Y, o_d[n] is the LSB of x_d[n], and T₁=−2, T₂=−1, T₃=0, T₄=1, T₅=2. The specific state transition matrices corresponding to the quantization block 202_d

$\begin{array}{cc}{A}_o=\left[\begin{array}{ccccc}0& 3/4& 0& 1/4& 0\\ 3/16& 0& 3/4& 0& 1/16\\ 0& 1/2& 0& 1/2& 0\\ 1/16& 0& 3/4& 0& 3/16\\ 0& 1/4& 0& 3/4& 0\end{array}\right]\mathrm{and}{A}_e=\left[\begin{array}{ccccc}1/4& 0& 3/4& 0& 0\\ 0& 5/8& 0& 3/8& 0\\ 1/8& 0& 3/4& 0& 1/8\\ 0& 3/8& 0& 5/8& 0\\ 0& 0& 3/4& 0& 1/4\end{array}\right].& \left(9\right)\end{array}$

Details on the derivation of these state transition matrices are provided below in the successive quantization detail section. These state transition matrices ensure that the (s[n])^p and (t[n])^p sequences are free of spurious tones because each is a random process whose autocorrelation function converges to a constant as its time spread increases. Furthermore, the PSD of s_d[n] has a zero at ω=0 and increases at 6 dB per octave as ω increases from zero. In this respect, the quantization noise shaping of this version of the successive requantizer is comparable to that of a first-order ΔΣ modulator.

Successive requantizers with higher than first-order quantization noise shaping can also be designed. For example, second-order quantization noise shaping can be achieved by quantization blocks that calculate s_d[n] as a function the running sum of t_d[n] in addition to t_d[n], a random sequence, and the parity of x_d[n]. However, the fractional-N PLL in this work is a phase noise cancelling fractional-N PLL, so higher than first-order shaping is not necessary because most of the quantization noise is removed prior to the loop filter via a DAC.

The running sums of each t_d,i[n] sequence, i.e.,

$\begin{array}{cc}{t}_{d,i}\left[n\right]=\sum _{k=0}^n{t}_{d,i-1}\left[k\right],t_{d,1}\left[n\right]=\sum _{k=0}^n{t}_d\left[n\right]& \left(10\right)\end{array}$

can be bounded for all n, and each t_d,i[n] has a smooth power spectral density that increases monotonically with frequency.

A tradeoff related to the quantization block 202_d in FIGS. 3A and 3B is that its reduced susceptibility to nonlinearity-induced spurious tones comes at the expense of increased quantization noise power. For example, if it is desired to have quantization noise with a first-order highpass spectral shape, but it is not necessary to prevent nonlinear distortion from inducing spurious tones in the quantization noise and its running sum, a quantization block that implements

$\begin{array}{cc}{s}_d\left[n\right]=\{\begin{array}{ccc}0,& x_d\left[n\right]=\mathrm{even}& \\ r_d\left[n\right],& x_d\left[n\right]=\mathrm{odd},& t_d\left[n-1\right]=0\\ 1,& x_d\left[n\right]=\mathrm{odd},& t_d\left[n-1\right]=-1\\ -1,& x_d\left[n\right]=\mathrm{odd},& t_d\left[n-1\right]=1\end{array}& \left(11\right)\end{array}$

can be used, where r_d[n] is an independent random sequence that takes on the values 1 and −1 with equal probability. In this case s_d[n] takes on values of −1, 0, and 1, whereas the t_d[n] generated by the quantization block of FIG. 3A takes on values of −1, 0, 1. Consequently, the power of the quantization noise from a quantization block based on (11) is significantly lower than that from the quantization block of FIG. 3A.

This example suggests what is believed to be a fundamental design tradeoff: reduced susceptibility to nonlinearity-induced spurious tones comes at the expense of increased quantization noise power. Increased quantization noise is unlikely to be a significant problem in phase noise cancelling fractional-N PLLs, but it is likely to be an issue in fractional-N PLLs that lack phase noise cancellation. Any particular design will benefit from analytical quantification of the tradeoff and its effect on the performance of a particular fractional-N PLL, and especially in the case of a PLL that lacks phase noise cancellation. Preferred embodiments of the invention use a phase noise cancelling PLL design, such as the design that is described in commonly owned U.S. patent application Ser. No. 12/352,293 (“the '293 application”), filed Jan. 12, 2009, and entitled Adaptive Phase Noise Cancellation for Fractional-N Phase Locked Loop (which application is incorporated by reference herein).

FIG. 4 is a block diagram of an experimental prototype phase noise cancelling fractional-N PLL prototype in accordance with the invention. The prototpye's reference frequency was 12 MHz, and its output frequency range covered the 2.4 GHz ISM band. The phase noise cancellation enables a loop bandwidth of 975 kHz which is close to the f_ref/10 loop bandwidth upper limit for stability.

The FIG. 4 fractional-N PLL includes the successive requantizer 200 to that is shown in FIGS. 2A-3B and the remainder of the circuit preferably conducts adaptive phase noise cancellation as conducted by the preferred embodiments in the '293 application. The FIG. 4 fractional-N PLL includes a sampled loop filter 402 that drives a VCO 404, with preferred embodiments using a sampled loop filter having two separate loop filter halves that drive separate half-sized parallel varactors in the VCO 404 as in the '293 application. An amplifier 406 uses a 12 MHz crystal and supplies the reference clock voltage to a phase frequency detector/charge pump 408. The timing of pulses in the reference voltage and pulses received from the divider 410 determines whether the PFD/Charge pump will produce a positive or negative pulse of current to increase or decrease the VCO frequency. The successive requantizer 200 acts as discussed above to provide a sequence of values that average to a desired fractional value. A noise cancellation path including a filter 412, dithered quantizer 414, segmented DEM 416 encoder and DAC bank 418, discrete-time integrates the quantization noise to obtain a digital sequence and transform it via the DAC 202 into a current pulse of fixed duration and variable amplitude that equals the quantization noise portion of the charge in the corresponding charge pump pulse. A prototype integrated circuit in accordance with FIG. 4 also included alternate circuitry that could be enabled or disabled, including a ΔΣ modulator which could be optionally be used instead of the successive requantizer and an offset pulse generator 420 that can be enabled or disabled. In the IC prototype, the loop filter's sampling could be enabled or disabled, and when it was disabled, the loop filter reduces to a conventional loop filter.

The divider 410 is similar to the divider in the '293 application, except with minor changes to provide timing signals that control the offset current generator 418 and open the loop filter switch each reference period. As described in '293 application, the necessary timing signals are obtained by a chain of flip-flops clocked at half the VCO frequency. The timing signal used to close the loop filter switch each reference period could similarly have been derived within the divider block, but an RC one-shot circuit, timed from an output of the divider block, with a nominal duration of 25 ns was used in the prototype integrated circuit instead for simplicity because the length of time the switch is left open is not critical. Provided the switch is open when the loop filter's input current is non-zero, the PLL dynamics are relatively insensitive to the length of time it is open. Testing of the prototype with the successive requantizer, offset pulse generator, and sampled loop filter enabled and α chosen such that αf_ref=50 kHz revealed fractional spurs well inside the 975 kHz loop bandwidth, but all below −70 dBc in power. Four copies of the IC were tested. Table 1 shows the worst-case measurements taken from the four ICs. An IC wiring mistake disabled the DAC calibration circuitry, so the measurements described above were made after a one-time manual adjustment of the DAC gain. To confirm the diagnosis of the mistake, it was corrected in one copy of the IC by FIB microsurgery, but with the anticipated side effect of a coupling path that increased the measured in-band phase noise, 3 MHz phase noise, and largest in-band fractional spur by 10 dB, 3 dB, and 3 dB, respectively, above those shown in Table 1.

Design Details
Technology	0.18 um 1P6M CMOS
Package and Die area	32 pin TQFN,
	2.2 mm × 2.2 mm
Vdd	1.8	V
Reference frequency	12	MHz
Output frequency	2.4-2.5	GHz
Measured loop bandwidth	975	kHz
Measured Current Consumption
VCO and Divider Buffer	5.9	mA	Core
Divider	7.3	mA	27.1 mA
Charge Pump, PFD, and Buffers	8.6	mA
Offset Current	0.6	mA
Digital	1.9	mA
DAC	2.8	mA
Bandgap Bias Generator	5.4	mA	9.8 mA
Crystal Buffer	2.7	mA
External Buffer	1.7	mA
Measured Integer-N Performance
Phase Noise at 100 kHz	−103	dBc/Hz
Phase Noise at 3 MHz	−125	dBc/Hz
Reference spur without sampling	−58	dBc
enabled
Reference spur with sampling	−70	dBc
enabled
Measured Fractional-N Performance
Phase Noise at 100 kHz	−98	dBc/Hz
Phase Noise at 3 MHz	−121	dBc/Hz
Worst case in-band fractional spur	−45	dBc
with ΔΣ modulator
Worst case in-band fractional spur	−64	dBc
with SR
Reference spur without sampling	−40	dBc
enabled
Reference spur with sampling	−70	dBc
enabled

Successive Requantization Proof

The restriction to first-order highpass shaped quantization noise discussed above still leaves flexibility in the design of the s_d[n] sequences. This flexibility is exploited to ensure that s^p[n] for p=1, 2, . . . , h, and t^p[n] for p=1, 2, . . . , h_t are free of spurious tones, where h_s and h_t are positive integers. By definition, if s^p[n] and t^p[n] contain spurious tones at a frequency ω_n, then Equations (2) and (3) above, respectively, are expected to be unbounded in probability at ω=ω_n as L>∞. Therefore, to establish that there are no spurious tones in either s^p[n] or t^p[n], it is sufficient to show that Equations (2) and (3) are bounded in probability for all |ω|≦π as L→∞. A spurious tone at ω=0 is just a constant offset. Many practical systems are able to tolerate, or compensate for this offset so this case is excluded from consideration. Theorems 1 and 2 below present sufficient conditions on the s_d[n] sequences for (2) and (3) to be bounded in probability for every L≧1 and 0<|ω|≦π, thereby ensuring the absence of spurious tones in s^p[n] and t^p[n]. First-order highpass quantization noise is achieved with quantization blocks that implement equation (11) above. Results imply that neither s_d[n] nor t_d[n] contain spurious tones. Therefore, s[n] and t[n] inherit these properties provided the r_d[n] sequences for d=0, . . . , K−1 are independent.

However, if the quantization noise or its running sum is subjected to non-linear distortion, spurious tones can be induced. The presence of spurious tones implies that subjecting t[n] to second-order distortion is sufficient to induce spurious tones even though t[n] is known to be free of spurious tones.

The spur generation mechanism can be understood by considering the first quantization block. Suppose the input to the successive requantizer is an odd-valued constant and t₀[n−1]=0 for some value of n. Then (11) implies that (s₀[n], s₀[n+1]) is either (−1, 1) or (1, −1) depending on the polarity of P₀[n]. It follows from (5) that (t₀[n], t₀[n+1]) is either (−1, 0) or (1, 0), and, by induction, t₀[n] has the form { . . . , 0, ±1, 0, ±1, 0, ±1, 0, . . . }. Therefore, t₀²[n] has the form { . . . , 0, 1, 0, 1, 0, 1, 0, . . . } which is periodic. A similar, but more involved analysis can be used to show that the t_d²[n] sequences for d>0 also contain periodic components. These periodic components cause the spurious tones.

For the purpose of illustrating the principle of successive requantization for tone free quantization sequences, the remainder of the description assumes that the input to the quantizer, x₀[n], is an integer-valued and deterministic sequence for n=0, 1, . . . , and that the successive requantizer is designed such that the following properties are satisfied:

• Property 1: x_d+1[n]=(s_d[n]+x_d[n])/2 is integer-valued for n=0, 1, . . . , and d=0, 1, . . . , K−1.
• Property 2: there exists a positive constant B such that |t_d[n]|<B, for n=0, 1, 2, . . . .
• Property 3: t_d[0]=0, andt_d[n]=f(t_d[n−1],r_d[n],o_d[n])  (12)
  • where r_d[n], d=0, 1, . . . , K−1, n=1, 2, . . . } is a set of independent identically distributed (iid) random variables, and

$\begin{array}{cc}{o}_d\left[n\right]=x_d\left[n\right]\mathrm{mod}2=\{\begin{array}{cc}1,& \mathrm{if}{x}_d\left[n\right]\mathrm{is}\mathrm{odd},\\ 0,& \mathrm{if}{x}_d\left[n\right]\mathrm{is}\mathrm{even},\end{array}& \left(13\right)\end{array}$

• • is called the parity sequence of the d^th quantization block. Furthermore, f(x, y, z) is a deterministic function which does not depend on n.

Property 2 guarantees first-order spectral shaping of the quantization error by ensuring that t_d[n] takes on a finite number of values for all n. However it need not be an optimal bound on the quantization error of the successive requantizer. Relaxing this bound, and hence incurring more quantization error power, permits the removal of spurious tones under non-linearities.

Property 1 and the assumption that x₀[n] is integer-valued imply that s_d[n] is an even integer when x_d[n] is even, and an odd integer otherwise. Therefore, (5) implies that t_d[n] is integer-valued, and Property 2 further implies that it is restricted to a finite set of values. Let T₁, T₂, . . . , T_N denote these values. Therefore, the function, ƒ, in Property 3 takes on values restricted to the set {T₁, T₂, . . . , T_N}.

It follows from Properties 1, 2, and 3 that x_d+1[n], s_d[n], and t_d[n], for d=0, 1, K−1,and n=1, 2, . . . , depend only on the set of iid random variables {r_d[n], d=0, 1, . . . , K−1, n=0, 1, 2, . . . } and the deterministic successive requantizer input sequence, {x₀[n], n=1, 2, . . . ,}. Therefore, the sample description space of the underlying probability space is the set of all possible values of the random variables {r_d[n], d=0, 1, . . . , K−1, and n=0, 1, 2, . . . }.

Equation (5) implies thats_d[n]=t_d[n]−t_d[n−1].  (14)

Therefore, it follows from Property 1 thatx_d[n]=t_d−1[n]−t_d−1[n−1]+x_d−1[n])/2,  (15)

for 1≦d<K. Recursively substituting (15) into itself and applying (13) yields

$\begin{array}{cc}{o}_d\left[n\right]=\frac{1}{2}\left[x_o\left[n\right]+\sum _{k=0}^{d-1}{2}^{-k}\left(t_k\left[n\right]-t_k\left[n-1\right]\right)\right]\mathrm{mod}2.& \left(16\right)\end{array}$

Recursively substituting (12) into itself implies that for any integer n>0,t_d[n]=g_n(r_d[n],r_d[n−1], . . . , r_d[1],o_d[n],o_d[n−1], . . . , o_d[1])  (17)

where g_n is a deterministic, memoryless function. Similarly, for any pair of integers n₂>n₁>0, recursively substituting (12) into itself m=n₂−n₁−1 times implies thatt_d[n₂]=h_m(t_d[n₁],r_d[n₁+1],r_d[n₁+2], . . . , r_d[n₂], o_d[n₁+1], o_d[n₁+2], . . . , o_d[n₂])  (18)

where h_m is a deterministic, memoryless function.

Repeatedly substituting (16) into (17) to eliminate the variables {o_d[n], . . . , o_d[1]} and then recursively substituting the result into itself to eliminate the variables {t_k[m], k=0, . . . , d−1, m=1, . . . , n} shows that t_d[n] is a random variable that depends only on x₀[n] (which is deterministic), and the random variables {r_k[m], k=0, 1, . . . , d, m=1, 2, . . . , n}. This in conjunction with (16) implies that o_d[n] is a random variable that depends only on x₀[n], and the random variables {r_k[m], k=0, 1, d−1, m=1, 2, n}. In particular, since the random sequence {o_d[n], n=0, 1, 2, . . . } does not depend on the random sequence {r_d[n], n=0, 1, 2, . . . } and since all the random variables {r_k[m] d=0, 1, . . . , K−1, n=0, 1, 2, . . . } are statistically independent by Property 3, it follows that {o_d[n], n=0, 1, 2, . . . } and {r_d[n], n=0, 1, 2, . . . } are statistically independent random sequences. By similar reasoning, the random variable t_d[n] is statistically independent of the random variables {r_d[m], m=n+1, n+2, . . . }

Hence, (18) implies that t_d[n₂] conditioned on the random variables t_d[n₁], o_d[n₁+1], o_d[n₁+2], o_d[n₂] is a function only of the statistically independent random variables r_d[n₁], r_d[n₁+1], . . . , r_d[n₂]. By definition, for i≠j the random variables {r_i[n₁], r_i[n₁+1], . . . , r_i[n₂]} are statistically independent of the random variables {r_j[n₁], r_j[n₁+1], . . . , r_j[n₂]}. Therefore, for i≠j the random variables t_i[n₂] and t_j[n₂] and conditioned on t_i[n₁], t_j[n₁], o_i[n₁+1], o_i[n₁+2], . . . , o_i[n₂], o_j[n₁+1], o_j[n₁+2], . . . , o_j[n₂] are statistically independent. Consequently, for any positive real numbers p₀, . . . , p_k−1,

$\begin{array}{cc}E\left[\prod _{j=0}^{K-1}{t}_j^{p_j}\left[n_2\right]❘t_d\left[n_1\right],o_d\left[n\right];d=0,\dots ,K-1,n=n_1+1,\dots ,n_2\right]=\prod _{j=0}^{K-1}E[t_j^{p_j}\left[n_2\right]❘t_d\left[n_1\right],o_d\left[n\right];d=0,\dots ,K-1,n=n_1+1,\dots ,n_2]=\prod _{j=0}^{K-1}E[t_j^{p_j}\left[n_2\right]❘t_j\left[n_1\right],o_j\left[n\right];n=n_1+1,\dots ,n_2],& \left(19\right)\end{array}$

where the second equality follows from (12) and the independence of the {r_d[n], n=1, 2, . . . ,} sequences for d=0, K−1. This implies that the pmf of the random variable t_i┌n₂┐ conditioned on t_i┌n₁┐, o_i┌n₁+1┌, o_i┌n₁+2┐, . . . , o_i┌n₂┐ is independent of any additional conditioning by and t_j[n₁], o_j[n₁+1], o_j[n₁1+2], . . . , o_j[n₂] for i≠j.

The statistical independence of o_d[n] and r_d[n] together with (12) imply that {t_d[n], n=0, 1, . . . } is a discrete-valued Markov random sequence conditioned on the sequence {o_d[n], n=0, 1, . . . }. Whenever x_d[n] is odd the one-step state transition matrix for t_d[n] is given byA_o=[r{t_d[n]=T_j|t_d[n−1]=T_i,o_d[n]=1}]_N×N.  (20)

Similarly, whenever x_d[n] is even the one-step state transition matrix for t_d[n] is given byA_e=[P{t_d[n]=T_j|t_d[n−1]=T_i,o_d[n]=0}]_N×N.  (21)

The function ƒ in Property 3 is independent of n and d, so neither matrix is a function of n and d.

Equation (14) implies that each possible value of s_d[n] is given by T_j−T₁ for some pair of integers i and j, 1≦i, j≦N, soP{s_d[n]=T_j−T_i|t_d[n−1]=T_i,o_d[n]=1}=P{t_d[n]=T_j|t_d[n−1]=T_i,o_d[n]=1}.  (22)

Given that t_d[n] is restricted to N possible values, s_d[n] is restricted to N′ possible values where N′≦N². With identical reasoning to that used to proceed from (15) to (19), it follows that

$\begin{array}{cc}E\left[\prod _{j=0}^{K-1}{s}_j^{p_j}\left[n_2\right]❘t_0\left[n_1\right],\dots ,t_{K-1}\left[n_1\right],o_d\left[n\right];d=0,\dots ,K-1,n=n_1+1,\dots ,n_2\right]=\prod _{j=0}^{K-1}E[s_j^{p_j}\left[n_2\right]❘t_j\left[n_1\right],o_j\left[n\right];n=n_1+1,\dots ,n_2].& \left(23\right)\end{array}$

Given that {t_d[n], n=0, 1, . . . } is a discrete-valued Markov random sequence conditioned on the sequence {o_d[n], n=0, 1, . . . }, the conditional probability mass function (pmf) of t_d[n₂] given t_d[n₁] and o_d[n] is equal to the conditional pmf of t_d[n₂] given t_d[n₁], t_d[n₁−1] and o_d[n]. Therefore, (14) implies that (23) is equivalent to

$\begin{array}{cc}E\left[\prod _{j=0}^{K-1}{s}_j^{p_j}\left[n_2\right]|s_0\left[n_1\right],\dots ,s_{K-1}\left[n_1\right],t_0\left[n_1\right],\dots ,t_{K-1}\left[n_1\right],o_d\left[n\right];d=0,\dots ,K-1,n=n_1+1,\dots ,n_2\right]=\prod _{j=0}^{K-1}E\left[s_j^{p_j}\left[n_2\right]|t_j\left[n_1\right],o_j\left[n\right];n=n_1+1,\dots ,n_2\right]& \left(24\right)\end{array}$

The following definitions are used by the theorems presented below. In analogy to the matrices A_o and A_e, letS_o=[P{s_d[n]=S_j|t_d[n−1]=T_i,o_d[n]=1}]_N×N′,  (25)andS_e[P{s_d[n]−S_j|t_d[n−1]=T_i,o_d[n]=0}]_N×N′,  (26)where {Si, 1≦i≦N′} is the set of all possible values of sd[n]. Property 3 ensures that neither matrix is a function of n and d. It follows from (22) that each non-zero element of So or Se is equal to an element in Ao or Ae, respectively. For example, if Sk=Tj−Ti, then the element in the ith row and kth column of S_o is equal to the element in the ith row and jth column of A_o. In this fashion, once A_o and A_e are known, S_o and S_e can be deduced.

Let

$\begin{array}{cc}1\stackrel{\Delta }{=}\left[\begin{array}{c}1\\ ⋮\\ 1\end{array}\right],t^{\left(p\right)}\stackrel{\Delta }{=}\left[\begin{array}{c}{\left(T_1\right)}^p\\ ⋮\\ {\left(T_N\right)}^p\end{array}\right],\mathrm{and}{s}^{\left(p\right)}\stackrel{\Delta }{=}\left[\begin{array}{c}{\left(S_1\right)}^p\\ ⋮\\ {\left(S_{N^{\prime }}\right)}^p\end{array}\right]& \left(27\right)\end{array}$

Suppose a sequence of vectors, b[n]=[b₁[n], . . . , b_N[n]]^T converges to a constant vector, b1, as n→∞. Then the convergence is said to be exponential if there exist constants C≧0 and 0≦α<1 such that|b_i[n]−b|≦Cαⁿ  (28)

for all 1≦i≦N and n≧0.

Theorem 1: Suppose that the state transition matrices A_e and A_o satisfyA_eA_o=A_oA_e,  (29)

and there exists an integer h_t≧1 such that for each positive integer p≦h_t

$\begin{array}{cc}\underset{n\to \infty }{\lim }{A}_e^n{t}^{\left(p\right)}=b_{p}1,\mathrm{and}\underset{n\to \infty }{\lim }{A}_o^n{t}^{\left(p\right)}=b_{p}1& \left(30\right)\end{array}$

where b_p is a constant and the convergence of both vectors is exponential. Then for every L≧1,E[I_t^p_,L^(ω)]≦C(ω)<∞  (31)

for each 0<|ω|≦π. Moreover, the bound C(ω), which is independent of L, is uniform in ω for all 0<ε<|ω|≦π.

By Markov's Inequality, this immediately leads to:

Corollary 1: Under the assumptions of Theorem 1, I_t^p_,L^(∞) is bounded in probability for all L≧1 and for each ω satisfying 0<|ω|≦π.

Proof of Theorem 1: The expectation of I_t^p_,L^(ω) can be expressed as

$\hspace{1em}\begin{array}{cc}\begin{array}{c}E\left[I_{tP,L}\left(\omega \right)\right]=\frac{1}{L}\sum _{n_1=0}^{L-1}\sum _{n_2=0}^{L-1}E\left[t^p\left[n_1\right]t^p\left[n_2\right]\right]e^{-\mathrm{j\omega }\left(n_1-n_2\right)}\\ =\frac{1}{L}\sum _{n_1=0}^{L-1}E\left[t^{2p}\left[n\right]\right]+\\ \frac{1}{L}\sum _{\underset{n_1\ne n_2}{n_1=0}}^{L-1}\sum _{n_2=0}^{L-1}E\left[t^p\left[n_1\right]t^p\left[n_2\right]\right]e^{-\mathrm{j\omega }\left(n_1-n_2\right)}\stackrel{\Delta }{=}{J}_1+J_2.\end{array}& \left(32\right)\end{array}$

The notation above means that J₁ and J₂ are defined as the first and second terms, respectively, to the left of the symbol. Property 2 states that |t_d[n]|≦B, so it follows from (6) that t[n]≦B₁ for some finite constant B₁. Therefore, J₁≦B₁^2p. The crux of the proof is showing that there exists a constant C_tp, positive constants D₁, D₂, and a constant 0<α<1 such that for n₁≠n₂ |E[t^p[n₁]t^p[n₂]−C_tp]|≦D₁α^|n2−n1|+D₂α¹¹  (33)

The proof of (33) is outlined in Lemma 1 in the proof of Lemmas given below. Here (33) is used to complete the proof of the theorem. From (32), J₂ can be expressed as

$\begin{array}{cc}{J}_2=\frac{1}{L}\sum _{\underset{n_1\ne n_2}{n_1=0}}^{L-1}\sum _{n_2=0}^{L-1}\left(E\left[t^p\left[n_1\right]t^p\left[n_2\right]-C_{t^p}\right]\right)e^{-\mathrm{j\omega }\left(n_1-n_2\right)}+\frac{1}{L}\sum _{\underset{n_1\ne n_2}{n_1=0}}^{L-1}\sum _{n_2=0}^{L-1}{C}_{t^p}{e}^{-\mathrm{j\omega }\left(n_1-n_2\right)}\stackrel{\Delta }{=}{J}_{2,1}+J_{2,2}.& \left(34\right)\end{array}$

From (33) it is seen that

$\begin{array}{cc}J_{2,1}\le \frac{1}{L}\sum _{\underset{n_1\ne n_2}{n_1=0}}^{L-1}\sum _{n_2=0}^{L-1}\left(D_1{\alpha }^{n_1-n_2}+D_2{\alpha }^{n_1}\right)\le \frac{D_1}{L}\sum _{n_1=0}^{L-1}\sum _{n_2=0}^{L-1}{\alpha }^{n_1-n_2}+D_2\sum _{n_1=0}^{L-1}{\alpha }^{n_1}\le 2\left(D_1+D_2\right)\frac{1-{\alpha }^L}{1-\alpha }\le 2\left(D_1+D_2\right)\frac{1}{1-\alpha }.& \left(35\right)\end{array}$

and the bound is independent of L. Similarly, J_2,2 can be bounded by

$\begin{array}{cc}J_{2,2}\le \frac{C_{t^p}}{L}{\sum _{n=0}^{L-1}{e}^{-\mathrm{j\omega }n}}^2-L\le \frac{C_{t^p}}{L}{\frac{1-e^{-\mathrm{j\omega }L}}{1-e^{-\mathrm{j\omega }}}}^2-L=\frac{C_{t^p}}{L}{\frac{\sin \left(\omega L/2\right)}{\sin \left(\omega /2\right)}}^2-L·\le C_{t^p}\left(1+\frac{1}{{\sin }^2\left(\omega /2\right)}\right)& \left(36\right)\end{array}$ which is finite, independent of L, for each ω satisfying 0<ω|≦π; the bound is uniform for all ω satisfying 0<ε<|ω|≦π since sin(ω/2)>sin(ε/2). The result of the theorem then follows from (32) through (36).

Theorem 2: Suppose that the state transition matrices A_e and A_o satisfyA_eA_o=A_oA_e,  (37)

and there exists an integer h_s≧1 such that for each positive integer p≦h_s, the sequence transition matrices S_e and S_o satisfy

$\hspace{1em}\begin{array}{cc}\begin{array}{c}\underset{n\to \infty }{\lim }{A}_e^n{S}_e{s}^{\left(p\right)}=\underset{n\to \infty }{\lim }{A}_e^n{S}_o{s}^{\left(p\right)}\\ =\underset{n\to \infty }{\lim }{A}_o^n{S}_e{s}^{\left(p\right)}\\ =\underset{n\to \infty }{\lim }{A}_o^n{S}_o{s}^{\left(p\right)}\\ =c_{p}1,\end{array}& \left(38\right)\end{array}$ where cp is a constant and the convergence of all vectors are exponential. Then for every L≧1,E[I_sp,L(ω)]≦D(ω)<∞  (39)

for each 0<<|ω|≦π. Moreover, the bound D(ω), which is independent of L, is uniform in ω for all 0<ε<|ω|<π.

By Markov's Inequality, this immediately leads to,

Corollary 2: Under the assumptions of Theorem 2, I_sp,L(ω) is bounded in probability for all L≧1 and for each ω satisfying 0<|ω|≦π.

Proof of Theorem 2: The proof is identical to that of Theorem 1. Replacing t^p[n₁] and t^p[n₂] with s^p[n₁] and s^p[n₂] respectively, the crux of the proof is showing that there exists a constant C_sp, positive constants E₁, E₂, and a constant 0<β<1 such that for n₁≠n₂ |E[s^p[n₁]s^p[n₂]−C_sp]|≦+E_tβ^|n2−n1|+E₂βⁿ¹,  (40)With (40) proven in Lemma 2 below, the remainder of the proof follows directly from Theorem 1.

Matrices A_e, A_o, S_e, and S_o which can be used with the successive requantizer to generate quantized sequences and satisfy the conditions of Theorems 1 and 2 for h_t=3 and h_s=5 are presented are provided as an example.

For a state t_d[n] whose possible values are {−2, −1, 0, 1, 2}, definet^(p)=[(−2)^p(−1)^p 0 1^p 2^p]^T  (41)

and the proposed state transition matrices as

$\begin{array}{cc}{A}_o=\left[\begin{array}{ccccc}0& 3/4& 0& 1/4& 0\\ 3/16& 0& 3/4& 0& 1/16\\ 0& 1/2& 0& 1/2& 0\\ 1/16& 0& 3/4& 0& 3/16\\ 0& 1/4& 0& 3/4& 0\end{array}\right];\mathrm{and}{A}_e=\left[\begin{array}{ccccc}1/4& 0& 3/4& 0& 0\\ 0& 5/8& 0& 3/8& 0\\ 1/8& 0& 3/4& 0& 1/8\\ 0& 3/8& 0& 5/8& 0\\ 0& 0& 3/4& 0& 1/4\end{array}\right].& \left(42\right)\end{array}$

From (14) all possible s_d[n] values are {−4, −3, −2, −1, 0, 1, 2, 3, 4}, and further defines^(p)=[(−4)^p (−3)^p (−2)^p (−1)^p 0 1^p 2^p 3^p 4^p]^T.  (43)

Applying (22) yields

$\begin{array}{cc}{S}_o=\left[\begin{array}{ccccccccc}0& 0& 0& 0& 0& 3/4& 0& 1/4& 0\\ 0& 0& 0& 3/16& 0& 3/4& 0& 1/16& 0\\ 0& 0& 0& 1/2& 0& 1/2& 0& 0& 0\\ 0& 1/16& 0& 3/4& 0& 3/16& 0& 0& 0\\ 0& 1/4& 0& 3/4& 0& 0& 0& 0& 0\end{array}\right],\mathrm{and}{S}_e=\left[\begin{array}{ccccccccc}0& 0& 0& 0& 1/4& 0& 3/4& 0& 0\\ 0& 0& 0& 0& 5/8& 0& 3/8& 0& 0\\ 0& 0& 1/8& 0& 3/4& 0& 1/8& 0& 0\\ 0& 0& 3/8& 0& 5/8& 0& 0& 0& 0\\ 0& 0& 3/4& 0& 1/4& 0& 0& 0& 0\end{array}\right].& \left(44\right)\end{array}$

Multiplying the matrices in either order yields

$\begin{array}{cc}{A}_e{A}_o=A_o{A}_e=\left[\begin{array}{ccccc}0& 9/16& 0& 7/16& 0\\ 9/64& 0& 3/4& 0& 7/64\\ 0& 1/2& 0& 1/2& 0\\ 7/64& 0& 3/4& 0& 9/64\\ 0& 7/16& 0& 9/16& 0\end{array}\right],& \left(45\right)\end{array}$ so the matrices commute. Direct computation reveals that the eigenvectors of both Ae and Ao are linearly independent, and therefore Ae and Ao are diagonalizable. Specifically, A_eⁿ=V_eλ_eⁿV_e⁻¹, where

$\begin{array}{cc}{\Lambda }_e^{″}=\left[\begin{array}{ccccc}1& 0& 0& 0& 0\\ 0& 1/4^{″}& 0& 0& 0\\ 0& 0& 0& 0& 0\\ 0& 0& 0& 1& 0\\ 0& 0& 0& 0& 1/4^{″}\end{array}\right],V_e=\left[\begin{array}{ccccc}1& 1& 1& 0& 0\\ 0& 0& 0& 1& -1\\ 1& 0& -1/3& 0& 0\\ 0& 0& 0& 1& 1\\ 1& -1& 1& 0& 0\end{array}\right],\mathrm{and}{V}_e^{-1}=\left[\begin{array}{ccccc}1/8& 0& 3/4& 0& 1/8\\ 1/2& 0& 0& 0& -1/2\\ 3/8& 0& -3/4& 0& 3/8\\ 0& 1/2& 0& 1/2& 0\\ 0& -1/2& 0& 1/2& 0\end{array}\right],& \left(46\right)\end{array}$

and A_oⁿ=V_oλ_oⁿV_o⁻¹, where

$\begin{array}{cc}{\Lambda }_o^n=\left[\begin{array}{ccccc}{\left(-1\right)}^n& 0& 0& 0& 0\\ 0& 1& 0& 0& 0\\ 0& 0& {\left(-1/4\right)}^n& 0& 0\\ 0& 0& 0& 1/4^n& 0\\ 0& 0& 0& 0& 0\end{array}\right],V_o=\left[\begin{array}{ccccc}1& 1& 1& -1& 1\\ -1& 1& -1/2& -1/2& 0\\ 1& 1& 0& 0& -1/3\\ -1& 1& 1/2& 1/2& 0\\ 1& 1& -1& 1& 1\end{array}\right],\mathrm{and}{V}_o^{-1}=\left[\begin{array}{ccccc}1/16& -1/4& 3/8& -1/4& 1/16\\ 1/16& 1/4& 3/8& 1/4& 1/16\\ 1/4& -1/2& 0& 1/2& -1/4\\ -1/4& -1/2& 0& 1/2& 1/4\\ 3/8& 0& -3/4& 0& 3/8\end{array}\right].& \left(47\right)\end{array}$

By inspection of (46), λ_eⁿ converges to

$\begin{array}{cc}{\Lambda }_{e,1}=\left[\begin{array}{ccccc}1& 0& 0& 0& 0\\ 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 0\\ 0& 0& 0& 1& 0\\ 0& 0& 0& 0& 0\end{array}\right],& \left(48\right)\end{array}$

The vector given by V_eλ_e,1V_e⁻¹t^(p) is equal to b_p1, where b_p=0, 1 and 0 for p=1, 2 and 3 respectively, which is of the form required by Theorem 1. To show exponential convergence, consider

$\begin{array}{cc}A_e^n{t}^{\left(p\right)}-b_{p}1=\left(A_e^n-V_e{\Lambda }_{e,1}{V}_e^{-1}\right)t^{\left(p\right)}\le A_e^n-V_e{\Lambda }_{e,1}{V}_e^{-1}t^{\left(p\right)},& \left(49\right)\end{array}$

where ∥ ∥ is the l₂ norm, and p=1, 2 or 3. Evaluating ∥t^(p)∥ for p=3, and ∥A_eⁿ−V_eλ_e,1V_e⁻¹∥ yields √{square root over (130)} and √{square root over (2)}(¼)ⁿ respectively therefore the right side of (49) is equal to√{square root over (260)}(¼)ⁿ  (50)

and therefore each element of the vector given by A_eⁿt^(p)−b_p1 converges exponentially to zero.

By inspection of (47), λ_oⁿ does not converge, however it is sufficient to show that the vector V_oλ_oV_o⁻¹t^(p) converges. Consider A_oⁿ=V_oζ_o,1ⁿV_o⁻¹+V_oλ_o,2ⁿV_o⁻¹ where

$\begin{array}{cc}{\Lambda }_{o,1}^n=\left[\begin{array}{ccccc}0& 0& 0& 0& 0\\ 0& 1& 0& 0& 0\\ 0& 0& {\left(-1/4\right)}^n& 0& 0\\ 0& 0& 0& 1/4^n& 0\\ 0& 0& 0& 0& 0\end{array}\right],\mathrm{and}{\Lambda }_{o,2}^n=\left[\begin{array}{ccccc}{\left(-1\right)}^n& 0& 0& 0& 0\\ 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 0\end{array}\right],& \left(51\right)\end{array}$

Multiplying V_oλ_o,2ⁿV_o⁻¹ by t^(p) for p=1, 2 or 3 results in a vector with all zero elements for all n≧1. Therefore, for all n≧1 and p=1, 2 or 3, A_oⁿt^(p)=V_oλ_o,1ⁿV_o⁻¹t^(p). By inspection, ζ_o,1ⁿ converges to

$\begin{array}{cc}{\Lambda }_{o,3}=\left[\begin{array}{ccccc}0& 0& 0& 0& 0\\ 0& 1& 0& 0& 0\\ 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 0\end{array}\right],& \left(52\right)\end{array}$

The vector given by V_oλ_o,3V_o⁻¹t^(p) is equal to b_p1, where b_p=0, 1 and 0 for p=1, 2 and 3 respectively. Replacing A_eⁿ,V_e, λ_e,1, and V_e⁻¹ in (49) with A_oⁿ,V_o, λ_o,3, and V_o⁻¹ respectively shows that ∥A_oⁿt^(p)−b_p1∥ converges exponentially to b_p1. Therefore the state transition matrices given by (42) satisfy the conditions of Theorem 1 for h_t=3.

Using the decomposition in (46) and (47) and the sequence transition matrices given by (44), it can be shown by direct computation that A_oⁿS_es^(p), A_oⁿS_os^(p), A_eⁿS_es^(p) and A_eⁿS_os^(p) converges to c_p1, where c_p=0, 1.5, 0, 6, and 0 for p=1, 2, 3, 4 and 5 respectively. Furthermore, the convergence of each vector at index n can be bounded using (49), replacing ∥t^(p)∥ alternately with ∥S_es^(p)∥ and ∥S_os^(p)∥, which implies that the convergence of A_oⁿS_es^(p), A_oⁿS_os^(p), A_eⁿS_es^(p) and A_eⁿS_os^(p) are exponential. Therefore, the matrices A., A_o, S_e, and S_o given in (42) and (44) also satisfy the conditions of Theorem 2 for h_s=5.

Proof of Lemmas 1 and 2

Lemma 1: Suppose the conditions of Theorem 1 are satisfied. Then there exists a constant C_tp, positive constants D₁, D₂, and a constant 0<α<1 such that for n₁≠n₂ |E[t^p[n₁]t^p[n₂]−C_tp]|≦D_tα|n^2−n1|+D₂αⁿ¹  (53)Proof of Lemma 1: To establish (54), it suffices to assume that n₂>n₁. Using (5), E[t^p[n₂]t^p[n₁]] can be expressed as

$\begin{array}{cc}E\left[t^p\left[n_2\right]t^p\left[n_1\right]\right]=\sum _{c_1=0}^{K-1}\dots \sum _{c_p=0}^{K-1}\sum _{d_1=0}^{K-1}\dots \sum _{d_p=0}^{K-1}{2}^{c_1+\dots +c_p+d_1+\dots +d_p}E\left[\prod _{i=1}^p{t}_{c_i}\left[n_2\right]\prod _{j=1}^p{t}_{d_j}\left[n_1\right]\right].& \left(54\right)\end{array}$

It is seen that the above expression is a finite sum of terms of the form

$\begin{array}{cc}Q\left(n_1,n_2\right)=E\left[\prod _{j=0}^{K-1}\left(t_j^{p_j}\left[n_2\right]t_j^{q_j}\left[n_1\right]\right)\right],& \left(55\right)\end{array}$

where p_j and q_j are positive integers less than or equal to p. It thus suffices to establish a bound for Q(n₁, n₂) of the form|Q(n₁,n₂)−C₃|≦C₁α^{n2 . . . n1}+C₂αⁿ¹.  (56)

The right side of (55) is computed by conditional expectation as follows

$\begin{array}{cc}Q\left(n_1,n_2\right)=E\left\{\prod _{i=0}^{K-1}{t}_i^{p_i}\left[n_1\right]E\left(\prod _{j=0}^{K-1}{t}_j^{q_j}\left[n_2\right]❘t_d\left[n_1\right],o_d\left[n\right],d=0,1,\dots ,K-1,n=n_1+1,\dots ,n_2\right)\right\}.& \left(57\right)\end{array}$

Substituting (19) into the inner conditional expectation of (57) yields

$\begin{array}{cc}Q\left(n_1,n_2\right)=E\left\{\prod _{j=0}^{K-1}\left(t_j^{p_j}\left[n_1\right]E\left[t_j^{q_j}\left[n_2\right]❘t_j\left[n_1\right],o_j\left[n\right],n=n_1+1,\dots ,n_2\right]\right)\right\}& \left(58\right)\end{array}$

Since {t_d[n], n=0, 1, . . . } is a Markov process for any given parity sequence, {o_d[n]=o_{d, n}, n=0, 1, . . . } where o_d,nε{0,1}, it follows from (20) and (21) that the m-step state transition matrix corresponding to t_d[n] from time n to time n+m can be written as

$\begin{array}{cc}{A}_d\left[n,m\right]=\prod _{k=n+1}^{n+m}\left[A_o{o}_{d,k}+A_e\left(1-o_{d,k}\right)\right],& \left(59\right)\end{array}$

where A_d[n, m] is an N×N matrix with elements of the formP{t_d[n+m]=T_i|t_d[n]=T_i,o_d[n+1]=o_d,n+1, o_d[n+2]=o_d,n+2, . . . , o_d[n+m]=o_d,n+m}.  (60)

Since o_d,n is either 1 or 0 for each n, (29) can be used to write (59) as

$\begin{array}{cc}{A}_d\left[n,m\right]=A_e^{y_m}{A}_o^{m-y_m}=A_o^{m-y_m}{A}_e^{y_m},\mathrm{where}{y}_m=\sum _{k=n+1}^{n+m}{o}_{d,k}.& \left(61\right)\end{array}$

By definition, y_m≧m/2 or m−y_m≧m/2 depending on the given parity sequence. It follows from the exponential convergence of (30) that there exists positive numbers C_p,e and C_p,o positive numbers α_p,e and α_p,o and less than unity such that each element ofA_e^ymt^(p)−b_p1  (62)

is less than C_p,eα_p,o^m/2 for y_m≧m/2, and each element ofA_o^m−ymt^(p)−b_p1  (63)

is less than C_p,oα_p,o^m/2 for m−y_m≧m/2.

The matrices A_o^m−ym and A_e^ym are stochastic matrices, so A_o^m−ym1=1, A_e^ym1=1 andA_o^m−ym(A_e^ymt⁽p)−b_p1)=A_o^m−ymA_e^ymt⁽p)−b_p1,  (64)A_e^ym(A_o^m−ymt⁽p)−b_p1)=A_e^ymA_o^m−ymt⁽p)−b_p1,  (65)

Since the elements of the vectors in (62) and (63) are exponentially bounded, the same must be true for the vectors in (64) and (65). From (61) it follows that the right side of either (64) or (65) is equal toA_d[n,m]t^(p)−b_p1.  (66)

Therefore, in general each element of (66) has a magnitude less than Cα^m/2 where C=max{C_p,e, C_p,o} and α=max {α_p,e, α_p,o}, which implies thatE[t_d^p[n+m]|t_d[n],o_d[n+j]=o_d,n+j, j=1, . . . , m]→b_p  (67)

as m→∞ uniformly in n where the convergence is also exponential. This result is independent of the given deterministic sequence {o_d,n, n=0, 1, . . . }, so it implies thatE[t_d^p[n+m]|t_d[n],o_d[n+j], j−1, . . . , m]→b_p  (68)

almost surely as m→∞ uniformly in n where the convergence is also exponential.

Thus, the inner conditional expectation in (58) converges exponentially to b_rj as n₂−n₁→∞ with probability one so that

$\begin{array}{cc}Q\left(n_1,n_2\right)\to \prod _{j=0}^{K-1}{b}_{q_j}E\left\{\prod _{i=0}^{K-1}{t}_i^{p_i}\left[n_1\right]\right\}.& \left(69\right)\end{array}$

More precisely, the exponential convergence of (69) implies that for every n₂>n₁ |E[t_j^qj[n₂]t_j[n₁],o_j[n],n=n₁+1, . . . , n₂]b_qj|≦C(q_j)αⁿ²⁻ⁿ¹.  (70)

with probability one where C(q_i) is a constant that depends on q_j. For every n₂>n₁

$\begin{array}{cc}Q\left(n_1,n_2\right)-\prod _{j=0}^{K-1}{b}_{q_j}E\left\{\prod _{i=0}^{K-1}{t}_i^{p_i}\left[n_1\right]\right\}\le E\left\{\prod _{j=0}^{K-1}{t_j\left[n_1\right]}^{p_j}E\left[t_j^{q_j}\left[n_2\right]❘t_j\left[n_1\right],o_j\left[n\right],n=n_1+1,\dots ,n_2\right]-b_{q_j}\right\}\le \prod _{j=0}^{K-1}C\left(q_j\right)B^{p_j}{\alpha }^{n_2-n_1}\stackrel{\Delta }{=}{C}_1{\alpha }^{n_2-n_1}& \left(71\right)\end{array}$

where B is given from Property 2. By similar reasoning, it can be established that

$\begin{array}{cc}E\left\{\prod _{j=0}^{K-1}{t}_j^{q_j}\left[n_1\right]\right\}-\prod _{j=0}^{K-1}{b}_{q_j}\le C_2{\alpha }^{n_1}& \left(72\right)\end{array}$

Hence, the above two bounds imply there exist positive constants C₁ and C₂ such that for all n₂>n₁

$\begin{array}{cc}Q\left(n_1,n_2\right)-\prod _{i=0}^{K-1}{b}_{p_i}\prod _{j=0}^{K-1}{b}_{q_j}\le Q\left(n_1,n_2\right)-E\left\{\prod _{i=0}^{K-1}{t}_i^{p_i}\left[n_1\right]\right\}\prod _{j=0}^{K-1}{b}_{q_j}+E\left\{\prod _{i=0}^{K-1}{t}_i^{p_i}\left[n_1\right]\right\}\prod _{j=0}^{K-1}{b}_{q_j}-\prod _{i=0}^{K-1}{b}_{p_i}\prod _{j=0}^{K-1}{b}_{q_j}.\le C_1{\alpha }^{n_2-n_1}+C_2{\alpha }^{n_1}.& \left(73\right)\end{array}$

Consequently, there exists a constant C₃ such that|Q(n₁,n₂)−C₃|≦C₁αⁿ²⁻ⁿ¹+C₂αⁿ¹  (74)

which is of the required form.

Lemma 2: Suppose the conditions of Theorem 2 are satisfied. Then there exists a constant C_sp, positive constants E₁, E₂, and a constant 0<β<1 such that for n₁≠n₂ |E[s^p[n₁]s^p[n₂]−C_sp]|≦E₁β^|n₂⁻ⁿ₁^|+E₂βⁿ¹  (75)

Proof of Lemma 2: The proof is similar to that of Lemma 1, so only the non-trivial differences with respect to the proof of Lemma 1 are presented.

Similarly to the proof of Lemma 1, it is necessary to show thatE[s_d^p[n+m]|t_d[n],o_d[n+j],j=1, . . . ,m]→c_p  (76)

almost surely as m→∞ uniformly in n where the convergence is also exponential. With this result and s_d[n], c_p, and (24) playing the roles of t_d[n], b_p, and (19) in the proof of Lemma 1, respectively, the proof of Lemma 2 is almost identical that of Lemma 1. Therefore, it is sufficient to prove (76).

Since the random variables t_d[n−1] and o_d[n] are statistically independent, for any given parity sequence, {o_d[n]=o_d,n, n=0, 1, . . . } where o_d,nε{0,1}, it follows from (25), (26), and (60) thatS_d[n,m+1]=A_d[n,m][S_oo_d,n+m+1+S_e(1−o_d,n+m+1)]  (77)

where S_d[n, m+1] is an N×N′ matrix with elements of the formP{s_d[n+m+1]=S_j|t_d[n]=T_i,o_d[n+1]=o_d,n+1, . . . , o_d[n+m+1]=o_d,n+m+1},  (78)

where i is the row index and j is the column index. By similar reasoning to that used in the proof of Lemma 1, (37) and (38) together imply that there exists a positive number D and a positive number β less than unity such that each element of the vectorS_d[n,m+1]s^(p)−c_p1,  (79)

has a magnitude less than D·β^m/2. Thus, (79) implies thatE[s_d^p[n+m]|t_d[n],o_d[n+j]=o_d,n+j, j=1, . . . , m]→c_p  (80)

as m→∞ uniformly in n where the convergence is also exponential. This result is independent of the given deterministic sequence {od,n, n=0, 1, . . . }, so it implies that (76) holds almost surely as m→∞ uniformly in n where the convergence is also exponential.

While specific embodiments of the present invention have been shown and described, it should be understood that other modifications, substitutions and alternatives are apparent to one of ordinary skill in the art. Such modifications, substitutions and alternatives can be made without departing from the spirit and scope of the invention, which should be determined from the appended claims.

Various features of the invention are set forth in the appended claims.

Claims

1. A successive requantizer device comprising: a plurality of quantization stages that each accept an input sequence and quantize the input sequence by a predetermined number of bits; whereineach quantization stage generates a quantization sequence sd[n], with the property that xd[n]+sd[n] is an even number for each n, where xd[n] is the quantization stage's input sequence; andeach quantization stage adds sd[n] to xd[n].

2. The device of claim 1, wherein each quantization stage discards a predetermined number of least significant bits to implement the quantization.

3. The device of claim 1, wherein each quantization stage simultaneously divides its input by 2b and quantizes the result by b bits every sample period

4. The device of claim 3, wherein each quantization stage simultaneously halves its input and quantizes the result by one bit every sample period.

5. The device of claim 1, wherein numbers within the successive requantizer device are integers with a two's-compliment binary number representation at each time n, and at each quantization stage sd[n] is chosen such that xd[n]+sd[n] does not exceed the range of a (K−d)-bit two's complement integer, and the parity of sd[n] is the same as that of xd[n], where K is determined by the number of quantization stages, the number of bits quantized per stage, and the logic which generates sd[n].

6. The device of claim 1, wherein that the running sum of each sd[n] sequence, i.e.,

7. The device of claim 6, wherein td[n] is a discrete-valued Markov random sequence conditioned on the parity of xd[n].

8. The device of claim 6, wherein td[n] is a random process whose autocorrelation function converges to a constant as its time spread increases.

9. The device of claim 6, wherein each quantization stage calculates sd[n] as a function the running sum of td[n] in addition to td[n], an independent random sequence, and the parity of xd[n].

10. The device of claim 6, wherein each quantization stage calculates sd[n] as a function of the previous value of td[n], the parity of the current value of xd[n], and the current value of an independent random sequence.

11. The device of claim 1, wherein each quantization stage implements

12. The device of claim 1, wherein that the running sums of each td,i[n] sequence, i.e.,

13. A fractional-N PLL device, comprising: a phase lock loop generating an output frequency from a voltage controlled oscillator based upon a low frequency reference voltage; anda successive requantizer according to claim 1 for providing a sequence of integers to control a divider in the phase locked loop to settle to a desired fractional value.

14. The fractional-N PLL device of claim 12, further comprising an adaptive phase noise cancellation path.

15. A successive requantizer device comprising: a plurality of quantization stages that each accept an input sequence and quantize the input sequence by a predetermined number of bits; whereineach quantization stage adds a quantization sequence to the quantization stage's input sequence and the quantization sequence is a discrete-valued Markov random sequence conditioned on the parity of the input sequence.

16. A successive requantizer device comprising: a plurality of quantization stages that each accept an input sequence and quantize the input sequence by a predetermined number of bits; whereineach quantization stage adds a quantization sequence to the quantization stage's input sequence and the quantization sequence is random process whose autocorrelation function converges to a constant as its time spread increases.