Active Components How to determine an effective damping factor for a third-order PLL A clever means for calculating (ζe) without using circuit simulations is presented. The advantage is that a relatively dense plot (55 points) of (ζe) vs. phase margin can be produced in a matter of seconds. By Ken Gentile U (s + α ) (1) s2 Where s is the complex frequency variable associated with the Laplace transform, α defines the zero of the open-loop response and K is the open-loop gain. This leads to a closed-loop response of the form: H CL ( s ) = H OL ( s ) 1+ H OL ( s ) = K (s + α ) s 2 + Ks + K α = ω K s+ n 2ζ s 2 + 2ζω n s + ω n2 (2) Note the introduction of two new variables: ζ (the damping factor) and ωn (the natural frequency of the loop). Both are expressed in terms of K and α, where and . The value of ζ correlates directly to the settling characteristics (transient response) of a second-order PLL, which is what makes it an attractive loop control parameter. A third-order PLL, on the other hand, has an open-loop response, which has the form: K (s + α ) H OL ( s ) = s 2 (s + β ) (3) In equation 3, α defines the zero and β the pole of the open-loop response. This leads to a closed-loop response of: H CL ( s ) = H OL ( s ) 1+ H OL ( s ) = K (s + α ) s3 + β s 2 + Ks + K α (4) Because the denominator takes the form of a cubic polynomial in s, the concept of a damping factor no longer makes sense. This is because a cubic has three factors, which, in general, would require three new variables defined in terms of K, α and β. With three variables defining the loop response, the transient behavior would be determined by the interaction of at least two of the variables. This precludes the use of a single variable as a gauge for the transient behavior of the loop. Hence, third-order PLLs are defined in terms of phase margin (φ) and open-loop bandwidth (ωc), which are related to α and β as given in equations 2 and 3. Even though third-order loops do not lend themselves to a damping factor parameter, Vaucher[1] showed that for a given value of φ an effective damping factor (ζe) can be obtained if one specifies a maximum amount of fractional settling error, ε. That is, ε can be expressed in terms of a specified transient frequency step size (fTRAN) and a specified maximum frequency error (fε) with respect to the final settling point such that (where fε < fTRAN). 32 706RFDF3.indd 32 1.4 1.2 1 0.8 0.6 �=30 �=50 �=70 0.4 0.2 0 0 0.5 1 1.5 2 Time (normalized) 2.5 3 Figure 1. Transient response of a third-order PLL. 0.6 0.4 Frequency (normalized) H OL ( s ) = K 1.6 Frequency (normalized) se of the damping factor (ζ) parameter as a gauge of the transient response of a second-order feedback loop is common in control theory. As such, it is a common practice to define the transient characteristics of a second-order phase-locked loop (PLL) in terms of ζ. The damping factor appears in the closed-loop response of a secondorder PLL, which may be derived from the open-loop response. In the s-domain, the open-loop response has the form: 0.2 0 -0.2 -0.4 �=30 �=50 �=70 -0.6 -0.8 -1 0 0.5 1 1.5 2 Time (normalized) 2.5 3 Figure 2. Transient frequency error as a function of time. Effective damping factor Figure 1 shows the transient behavior of a third-order PLL for three different values of φ. The plot is normalized to the frequency transient step size (fTRAN) on the vertical and to 1/fc (or 2π/ωc) on the horizontal. The most notable aspect of these curves is that φ has a direct impact www.rfdesign.com June 2007 6/21/2007 4:00:45 PM � �� �� ���� ���� ���� �� �� �� ����������������������������������� ����������������������������������� � �� �� �� �� �� ��� ��� ��� ��� �� �� �� �� �� �� ����������������� �� �� �� �� �� �� �� �� �� ��� ��� ��� ��� �� ���� ���� ���� �� �� �� �� �� �� �� ����������������� �� �� �� Figure 3. Logarithmic transient frequency error as a function of time. Figure 4. Approximating the slope of the envelope. on the overshoot and settling characteristics of the loop. So it would seem reasonable that some corollary could be drawn between φ in a third-order system and ζ in a second-order system. This was shown to be true according to Vaucher[1]. It is generally understood that ζ is a parameter related to the time required for a second-order PLL to settle to some acceptable level of frequency error following a transient frequency step. Since our goal is to relate φ to an “effective” ζ, it makes sense to view the transient step response in terms of frequency error relative to the final steady-state value. This is shown in Figure 2 for the same three values of φ. Note that the steady-state value corresponds to 0 on the vertical scale and the traces now display deviation from steady state as a function of time. Although helpful in visualizing the transient error, Figure 2 does not provide much insight into an analytical solution for relating ζ to φ. However, if the transient error is plotted on a log-scale, an interesting observation can be made. This is shown in Figure 3. Note that the horizontal axis has been extended, because plotting the transient error logarithmically makes it easier to view a much wider dynamic range of frequency error. Also, dashed lines have been added that indicate the slope of the envelope of each of the traces. The straight-line nature of the trace envelopes is an important observation. Notice that the slope of the envelope provides a linear relationship between the logarithmic frequency error and time. That is, given a value of φ (which defines a particular trace) and some specified maximum acceptable relative error threshold (in nepers), we see that the time required to reach the threshold level may be derived from the slope of the trace envelope. It appears that the slope of the envelope is the connection between φ and an “effective damping factor,” ζe. In fact, Vaucher[1] makes the argument that ζe can be defined as the inverse of the slope of the envelope, where the envelope is defined by observing the normalized logarithmic frequency error as a function of time. It is also interesting to observe that as φ increases from 30 to 50, the slope of the envelope becomes steeper. However, as φ increases from 50 to 70, the slope becomes shallower. This would imply that there may be some optimal value of φ that yields the quickest time to settle to a given error threshold level. In fact, this is shown to be true and is demonstrated at the conclusion of this article. generating the transient error data is relatively straightforward, but a computational method for identifying the envelope in order to determine the slope is not trivial. My solution to the latter problem is based on the following observation. Given a particular value of φ and specified frequency error threshold level, draw a line from the origin to the point on the frequency error trace where the frequency error first falls below the threshold level. This line is a fairly good approximation of the slope of the envelope. This is shown in Figure 4. In keeping with Figure 3, a dashed line indicates the slope of each trace envelope. The arrow-tipped lines indicate the aforementioned approximation. For this example, an arbitrary threshold level of -9.5 nepers was chosen. Notice that the slope of each arrowed line is a reasonable approximation to the slope of the associated envelope. Hence, it is reasonable that the slope of the arrowed line can be used to approximate ζe instead of the actual slope of the envelope. The reason for using this approximation is that determination of the slope of the arrowed line is a much more tractable problem than the determination of the actual slope of the envelope. Using this method for estimating the slope of the envelope, the details of the ζe computational process may be addressed as shown in Figure 5. The input parameters are phase margin, normalized logarithmic threshold level and lock time. Only the first two parameters are required to determine ζe. The lock time is only required if a calculation of the minimum necessary loop bandwidth is desired. The phase margin is used as the parameter to determine the K, α and β coefficients in the closed-loop frequency response as given by equation 4. The coefficients are calculated based on the methodology given in references 2 and 3. Normally, in addition to the desired phase margin, the open-loop bandwidth (ωc) is required to calculate the coefficients. However, the open-loop bandwidth simply scales all of the computations, so the analysis can be accomplished by normalizing the bandwidth to unity (fc = 1 Hz) and scaling the bandwidth dependent results by fc. The normalized values of K, α and β (as a function of φ) are given by: Computational algorithm To my knowledge, nowhere in the published literature on this subject is there a closed-form solution for relating φ to ζe. Hence, some method must be employed that can generate transient error data for a given φ. Once the transient error data is available, a method for determining the slope of the logarithmic error envelope must be resolved, as ζe is directly related to the slope of the envelope. The method for 34 706RFDF3.indd 34 K = (2π ) 2 1 + sin (φ ) 1 − sin (φ ) 1 − sin (φ ) α = 2π cos (φ ) cos (φ ) β = 2π 1 − sin (φ ) www.rfdesign.com June 2007 6/21/2007 4:00:49 PM ���������� ���������� ��������� ������� �������� �������� � ��������� ���� ������������ ��������� ���� ���� ���������� ���������� ���� ����� �������� �������� �������� ���� �������� ���������� ����������� ��������� ��� ���������� ����������� ����� �������� � ��������� ������� ������� ��������� �� �� ������� ��������� ������� ���� ��������� ������� ���� ���� ������ Figure 5. Computational process to determine ζe. The previous equations show that φ is an argument to several trigonometric functions. Normally, φ is specified in degrees, but it should be noted that most programming languages will require that φ be converted to radian units to properly compute the trigonometric functions. The closed-loop response, H(s), is computed per equation 4 using the above coefficients. For computational purposes, the Laplace variable, s, is replaced by jω (or j2πf). The reason for generating the closed-loop response is to ultimately derive the transient response, from which follows the error response and its associated envelope (as demonstrated in Figure 4). In order to produce an error response with sufficient resolution over a broad range of phase margin values it is necessary to compute the closed-loop response over a sufficiently wide frequency range. It was found that a frequency span covering 700 times the closed-loop bandwidth is sufficient for most cases. With the closedloop bandwidth normalized to 1 Hz this equates to 0 ≤ f ≤ 700. The closed-loop response is transformed into the time domain by means of an inverse FFT. The result is the impulse response of the closed-loop response. However, an inverse FFT applied directly to H(s) without modification will result in a complex impulse response. This is due to the fact that H(s) is usually expressed for positive frequencies. Since we know that the impulse response must be real (i.e., no imaginary numbers), H(s) must be modified to include the appropriate negative frequencies, as well. Doing so will cause the inverse FFT to produce an impulse response that contains only real values. Furthermore, in order to produce an impulse response with adequate time resolution it is necessary to ensure that H(s) contains enough frequency samples before invoking the inverse FFT. It was found that computing H(s) with 215 frequency points over the range -700 ≤ f ≤ 700 yields satisfactory results. The result is that the time resolution for the impulse response is 1/1400 in normalized units. In units of seconds, this equates to 1/(1400fc), where fc is the open-loop response in Hz. The step response is computed by convolving the impulse response with a step function. The step function is nothing more than a vector of ones that is the same length as the impulse response. This yields a step response that is normalized to unity, but it scales linearly with any arbitrary transient step size. The error response is calculated by subtracting 1 from normalized step response. The logarithmic error response is calculated from the error response. 36 706RFDF3.indd 36 However, since the error response contains positive and negative values, the logarithmic error response is computed using the absolute value of the error response. The result is a data set similar to that shown in Figure 3, but containing only a single trace associated with the specified value of φ. Finally, the normalized logarithmic error response is analyzed to find the point at which the data remains below the threshold level. The normalized logarithmic threshold level is specified in nepers. It represents the ratio of the absolute maximum allowable frequency error to the magnitude of the initial frequency step transient. For example, if the initial frequency step transient is 7 kHz and maximum allowable frequency error for declaring frequency lock is 5 Hz, then the threshold level is calculated as ln(5/7000) = -7.24 nepers. Given the normalized logarithmic threshold level (NTHRESH) in nepers, it is possible to find the normalized time point (tTHRESH) at which the data crosses permanently through NTHRESH. Then, the slope of one of the arrow-tipped lines in Figure 4 is calculated as: Since the effective damping factor (ζe) may be approximated as the inverse of the slope of the arrow-tipped line, then: If one wishes to know the minimum loop bandwidth (fc_min) required for a specified lock time, then the desired lock time (tLOCK) must be provided. With tLOCK specified in units of seconds, the minimum loop bandwidth (in hertz) is expressed as: Effective damping factor as a function of phase margin With the methodology outlined above it is possible to compute ζe over a range of φ values for a specified threshold level. This provides a means to build a plot of ζe vs. φ that can serve as a tool for identifying the effective damping factor associated with a particular phase margin www.rfdesign.com June 2007 6/21/2007 4:00:52 PM ��� � ��� � � ��� ������������������� ������������������������ ��� ��� ��� ��� � � � ��� ��� ��� � ��� ��� ��� ��� ��� �� � ��� ���������������������� ��� Figure 6. Effective damping factor vs. phase margin. �� �� ��� ��� �� ����� � � � � NTHRESH = -10 nepers � ��� tLOCK = 100µs ��� ��� ��� ���������������������� Figure 8. Minimum open-loop bandwidth vs. phase margin. The results given in reference 1 demonstrate that an “effective” damping factor (ζe) can be determined for a third-order PLL that is somewhat analogous to the commonly used damping factor parameter (ζ) in second-order PLLs. In reference 1, circuit simulations were used to generate the data for the time domain transient waveforms required to determine ζe. For a given phase margin (φ), multiple simulations were executed with various combinations of the transient step size and closed-loop bandwidth (ωc) and the results averaged to arrive at a mean value of ζe for the specified φ. The technique described builds on the work given in reference 1 by eliminating the need to run multiple circuit simulations. Instead, the time domain waveforms are generated from the closed-loop transfer function after determining the necessary coefficients as described in references 2 and 3. This technique allows the analysis to be normalized to the closed-loop bandwidth and the transient frequency step size, which eliminates the need for multiple simulations and averaging. The caveat is the introduction of the small error associated with approximating the slope of the logarithmic error envelope rather than using the actual slope. However, the relatively small error introduced by this approximation is worth the dramatic reduction in processing time compared to the methodology used in reference 1. In fact, the procedure outlined above that produced Figure 6 through Figure 8 was executed in less than 10 seconds using a PC with a 2.4 GHz dual-core processor. RFD 706RFDF3.indd 38 ��� �� ∆φ = 0.5º 38 ��� �� φmax = 80º Conclusion ��� ���������������������� Figure 7. Settling time vs. phase margin. ������� value (based on a specific threshold level). Since the above procedure yields tTHRESH and fc_min, then it is a simple matter to also generate plots for tTHRESH vs. φ and fc_min vs. φ. Figure 6 through Figure 8 are plots that were generated using MATLAB to execute the procedure outlined above. The dashed trace indicates the raw data generated by the procedure. The solid trace is data after smoothing. The ripple that appears in the raw data can be attributed to the approximation of the slope of the envelope (the arrowed lines in Figure 4) rather than the true slope (the dashed lines in Figure 4). The results agree reasonably well with those presented in reference 1. The slight deviation in the values shown in the plots here with respect to those shown in reference 1 can be attributed to the same slope approximation error mentioned above. The plots here were generated with the following parameters: φmin = 15º ��� References 1. Vaucher, C. S., “An Adaptive PLL Tuning System Architecture Combining High Spectral Purity and Fast Settling Time,” IEEE Journal of Solid-State Circuits, vol. 35, No. 4, April 2000. 2. Hawkins, D. W., “Digital Phase-Locked Loop (PLL)-Based Frequency Synthesizers: Theory and Analysis,” June 18, 1999, World Wide Web. 3. Keese, W. O., “An Analysis and Performance Evaluation of a Passive Filter Design Technique for Charge Pump Phase-Locked Loops,” National Semiconductor Application Note 1001, May 1996. 4. Wolaver, D. H., Phase-Locked Loop Circuit Design, Prentice Hall, 1991. ABOUT THE AUTHOR Ken Gentile is a system design engineer for the Clock and Signal Synthesis Products Group at Analog Devices, Greensboro, NC. His specialties are the application of digital signal-processing techniques in communications systems and analog filter design. He holds a B.S.E.E. degree from North Carolina State University. www.rfdesign.com June 2007 6/21/2007 4:00:55 PM