A New, Continuous-Time Model for Current-Mode Control R.B. Ridley ABSTRACT The accuracy of sampled-data modeling is combined with the simplicity of standard pole-zero represen- tation to give a new carent-mode contra’ model, accurate 10 half the switching frequency. All of the small-signal characteristics of current-mode control are predicted, including low-frequency effects and high-frequency subharmonic oscillation which can ‘occur even at duty cycles of less than 0.5. The best representation for the control-to-output transfer function is shown to be third-order. Model pred- ictions are confirmed with measwements om a buck converter. 1. Introduction Current-mode control has been Used in switching power supplies for several years. Numerous at- ‘tempts have been made to characterize this control system with small-signal models, all with limited degrees of accuracy or usefulness. Some continuous-time models [1-3] provide low- models for the system, but they need to address the well-known phenomenon of current- oop instability as a separate issue. Other models [4] have attempted to explain this instability through a modulator gain model, but predictions are not confirmed by measurements. “Exact discrete-time and sampled-data models [5,6] can accurately pre- dict responses, but they provide very little design insight due to their complex formulations. In this paper, the accuracy of sampled-data model- ing is combined with the simplicity of the three terminal PWM switch |7] to provide # complete model which accurately predicts characteristics, from dc to half the switching frequency. It is shown that an approximation to the sampled-data model can provide a model with a finite number of poles. Feedforward and feedback pain terms from input voltage and output voltage are derived, and are shown to be significantly different from earlier der- ivations [1]. Experimental results are presented to confirm the validity of the new model. II. Voltage-Mode Contro} Model Recent advances in analysis [7] have provided a flexible small-signal model which recnains invariant in the different PWM converters and is easy 10 use. Fig. 1 shows the invariant PWM three-terminal switch model developed in [7]. The voltage source is determined by the steady-state de voltage across the active and passive terminals, and by the duty cycle of the power stage. The current source is de- termined by the steady-state dc current, /,, out of the common terminal of the three-terminal model. ‘These quantities will depend upon the input volt- age, output voltage, and steady-state inductor cur- rent of the converter in which the model is placed. Fig. 2 shows the PWM switch model configured for the buck, boost, and flyback circuits, Point-by- nt substitution of the model of Fig. | into the BWM block gives the smallsignal model of the power stage. Fig: PWM Thre-Terminal SimalSignl Switch Model, This Baal an be ose or al eesrunsh PWM conte spare yi ie comunes omancon moses The torr ues Tr ed bite Scrmined te de opening sons oPiicfomt age (© 1989 PCIM. Reprinted, with permission, from Proceedings of the Power Conversion and Intelgence Motion; Long Beach, CA, ‘October 16619, 1989; pp. 455.066. 321” Pw P + R 7 e 5 R Patt Fig. 2: PWM Converters with Switch Model Inserted. Point: bypoint substitution of the model of Fig. 1 gives the complete power stage small-signal mode for te (a) buck, (b) boost, and (6 fpback converters. ‘The duty cycle, d, for control of the converter is typically generated with a control voltage and a reference ramp clocked at the desired switching fre- quency. A naturally-sampled duty cycle modulator is shown in Fig. 3. A sawtooth ramp of slope S, intersects a control voltage, ¥. , to produce the control parameter duty cycle, d. The small-signal model for this modulator has been found [8] to be: 1 STs Fra wo he where T, is the switching period. The model in Fig. 2, when combined with the duty-cycle modulator gain, Fy, , gives small-signal transfer functions which can be shown experimentally to be accurate to half the switching frequency. Fig. 3: Naturaly-Sampled Duty-cyele Medatater. The paturally-sampled modulator for singk-loop control, consisting of a satiooth ramp waveform intersecting 2 voltage reference, Fe modeled with a simple gain block, Fy IIL Current-Mode Control Model Fig. 4 shows a current-mode control modulation scheme. A constant-frequency clock initiates the on time of the switch, and the modulator ramp inter- sects a threshold to turn the switch off. The current ‘waveform provides the ramp of the modulator, and is supplemented with an external ramp to provide design flexibility and stabilize the current feedback loop [1-4 If the combined modulator slope, given, by the sum of the extemal ramp and the current ramp, is the same as for voltage mode control, the ower stage model and modulator gain for current- ‘mode control remain the same, and added feedback terms should account for all the phenomena of current-mode control. The modulator gain of the circuit is then a mS, + SOT; 4 57s @ OST where malt ® and 5, is the on-time slope of the current-sense waveform. This modulator gain is different from that found in [1} and (4), resulting in very significant changes to the model. 322ae Fig. 4 Current-Mode Control Modolator. ‘The ramp of the sted current signa ie summed wth a sawtooth ramp, and Compared witha Felerencevoliage to conta the Uy eee. POWER STAGE MODEL Fig. §: Complete Small-Signal Medel for Current-Mode Control ‘The appropriate power eeage mode! from Fig. ? ix used, and & is equal tothe indctor corent. Feedback path, ky and ba a8 ‘Gealed By closing the curren feedback path. Gain’ block 7740) represents the stipling acon of the converter. The external ‘aimp added to the eeu only affec the modulator in, Fig. 5 shows a complete block diagram for PWM Gxeuits with curent-mode control. The power stage model temains the same as that in Fig. 2, and gain blocks R, and H,(9) represent the current feed back. R; is the linear gain of the current-sense net- work, and H,(8) represents the sampling action of the curent-mode control. As will be seen later, proper approximation for this sampling block re~ sults in a powerful new small-signal model. Gains iy and k, provide feedforward of the input voltage ahd feedback of the output voltage, respectively. ‘These gain paths are created by feedback of the inductor current, the slope of which depends upon the input and output voltages of the converter. IV. Sampled-Data Current-Feedback Transfer Function The power stage mode! of Fig. 2 provides accurate transfer functions with voltage-mode control with- out the need for any discrete-time or sampled-data modeling. However, current-mode control exhibits characteristics which can only be explained with discrete-time modeling. It is not necessary to at- tempt to model the complete power stage with discrete-time or sampled-data analysis. Only the current-sampling function needs to be modeled, and then converted into continuous-time representation and combined with the rest of the power stage and feedback models, For the discrete-time analysis, input voltage and output voltage are kept constant, and the control-to-inductor current transfer func- tion is derived, with oke curent-feedback loop closed. The gain Hs) can then be found by inser ing appropriate values into the current foop of Fig. 5. kat © Fig, 6: Current-Mode Control Medulator Waveforms. Condi. are shown, The senged [ote fort change mental shag torent i cal io te dts feeaceetwet he enya Seber li ae Sat sapg ow aed net. ig. 6) plots he eae eran re Steauy-ate curt ewterm, snd Pg. 9) show he ca ‘Seprheocae srple ana Rid te 323Fig. 6 shows the sensed inductor curent waveforms, scaled by the feedback resistor R; with fixed input and output voltage for constant- frequency control, with the clock initiating the on- time of the power switch. ‘The solid line sepresents the steady-state condition, and the dashed line shows the perturbation waveforms. Fig. 6(b) shows the exact instantaneous perturba- ‘ion of the inductor current from the steady-state condition, and Fig. 6{c) shows the approximate equivalent sample-and-hold system waveforms. The only difference in the actual perturbation and the equivalent sample-and-hold system is a slight vari- ‘ation in the sampling instant, and a finite slope in the exact current. These differences are minor, and the constant-frequency current-mode control sys- tem can be considered a sample-and-hold system with the sampling instant occurring at the inter- seation of the current signal and the reference ‘waveform. ‘The discrete-time equation describing the eq ‘alent sample-and-bold function is e+ dae + d+ aery where SS St 5e @ ‘The z-transform of Eq. (4) is given by 4 4 2 Ny Tog tage © Notice that, with S, < Sy and no extemal ramp, the value of a is greater than unity, and the diserete- time system has a pole outside the unit circle. This explains the system instability at doty cycles greater than 0.5. However, instability can occur at duty cycles lower than 0.5, and this will be apparent when the complete system model is developed. ‘The transformation of a first-order sample-and-hold system from discrete-time into continuous time is analyzed thoroughly in [9]. The continuous time representation of the sample-and-hold circuit can ‘be found from the z-transform expression by using the substitution z=e% , and multiplying by (l-e*7))sT, . ‘The continuous-time represen tation of Eq. (6) is then given by: h@ _1 dee eat O50" R OT, hae Eq. (7) can be used with the model of Fig. $to find the open-loop sampling gain H,(s). With input and ‘output voltages constant, the gain terms ky and k, are not needed. All of the other gains of Fig. 5 are defined, and the power stage model is the same as for voltage-mode conteal. Hg) = Woy (8) Jt can be shown that this sampling gain is invariant forall converters using constant frequency, constant on-time, of constant offtime control. This invari ant equation can now be approximated to give a simple model. Y. Continuous-Time Approximation to Sampled-Data Model The sampled-data model has been realized before {5} but it has never been exploited to its full poten tial. The exact contisuous-time model of Eq, (8) has an infinite number of poles and zeros [5]. $i such a representation is not useful for design and analysis, applications of the sampled-data model have been limited. Attempts to model all of the poles of the sampled- data model in Eq, (8) are neither necessary nor useful. In fact, it is only necessary to accurately model the sampled-data expression. up to half the switching frequency. A complex pair of RHP zeros provides an accurate representation of the transfer function H,(s). This second-order model of the sampled-data system can be chosen to match the exact equation at the lower and upper limits of the frequency range of interest, from de to half the switching frequency. The transfer function of this second-order model is 2 a ro) where (to) and ay Fig. 7 plots the exact sampled-data model of Eq. (8) and the approximate second-order model of Eq. (9). It can be seen that the approximate mode! is exact at dc and half the switching frequency, and deviates by less than 0.2 dB and 3 degrees at fre- ‘quencies in between. Notice that this model gives a itation which has two poles and three zeros, with two of the zeros in the complex right-half plane. The reason for this apparent anomaly is the choice of a model which is good only to half the switching frequency. Ifthe model is extended to higher frequencies, more poles 324will be needed for accurate modeling, and the number of zeros will not be greater than the num- ber of poles. ‘The extta zeros in the current feed- back Ieop will cause additional poles in the closed-loop transfer functions, leading to the signif- icant differences and usefulness of the new model. VI. Complete Continuous-Time Model ‘The new curent-mode control model of Fig4 can now be completed with the derivation of feedforward gain k, and feedback gain k, . The av- erage inductor curent of the cigcuit, used in the power stage model, is related to the instantaneous current, used in the modulator, through the current ipple. “The current ripple is affected by Goth the input and output voltages of the citcuit, and the feedforward and feedback gains are used to model this dependence. Referring to the steady-state waveforms of Fig. 6, the describing function for the inductor current is given by S/D'Ts z The quantity denotes the average value of inductor cutrent “under steady-state conditions. This equation can be perturbed, assuming V, is constant, to obtain the dependence of inductor current On input voltage. ‘The gains of Fig. 5 can then be substituted to give the desired results for the gain ky for each of the converters. Eq. (12) can also be peiturbed, with V% constant, to obtain the de- pendence of inductof current on output voltage. ‘The gains of Fig. 5 can then be substituted to give the desired values of k, for each of the converters. The values of both fy and , are presented in Table 1. With the derivations of ky and &,, the new current-mode control modet Is complete. This Ri = V,~DTS,- (2) model can now be used for accurate analysis and design. Sane ‘Qvasratic — es ahssesas 3 rao fe Fig, 7: Exact Samplet-Data Model and. Quadtatc Approx tration ‘The ec a approximate expresons match exactly Bde and fi the swicingtequenay, and dif by ess than du and’ agree a a requence’ i berween, The current oop ey scaly nen, foe by ate ich has @ pu of complex RHP efor at hal the Swhchlogeaueey, adtceTo the eral pole and sro of Secure oor, TABLE t Summary of Gain Parameters of Fig. 5 325VII. Transfer Functions of New Model ‘A. Buck Converter Example The benefits and features of the new small-signal model are clearly demonstrated with an example. The buck converter shows some of the most inter- esting characteristics with current-mode control, so this converter was modeled with the following pa- rameters: Vea 1v v Le375pH = C=400uF = R ReaMrQ — R=0330 T= Dus The small-signal parameters of the three-terminal fitch and current-mode model can be calculated vy FBH5A Vg=¥=V D=045 Ky= 0.0619 &, = 0.088 ‘The small-signal model of Fig. 4 was built with a PSpice [12] file, using the above circuit parameters. ‘The second-order approximation of the sampling sampling block, 1/6), in Ea. (9), was easly built in PSpice using a simple operational amplifier net- work. The PSpice model was used to generate all of the transfer functions in this section of the paper. B. Current Loop Gain, ‘The first transfer fonction of interest is the loop. gain measured at the output of the duty-cycle modulator with the current loop feedback closed. ‘This loop gain determines the stability of the cur- sent feedback, and will show the cause of subhar- monic oscillation. Fig. & shows a plot of the current-loop gain with different values of extemal ramp. For the case with no extemal ramp (m_= 1), it cam, be seen that there is very little gain ‘or phase margin in this loop. If the duty cycle in- creases further, the gain increases and the system becomes unstable at D= 0.5. The shape of the gain and phase curves do not change with added external ramp. Even with low current feedback (m= 8) , there is significant gain at the filter resonant fte- quency, and the other transfer functions of the converter will be considerably altered from obtained for voltage-mode control. Tt can be shown thet the gain block k, has little ef- fect on the current-loop gain. ignoring this gain, the approximate current loop gain of the buck com verter is L__1+scr TO BDitm ay OD The denominator A(s) is the normal power stage transfer function denominator given by My + aa te ay where a5 and a6) “T[Era] 4+cr] This curent-loop gain expression differs. signif- icantly in form from the averaged models (1-4]. The complex RHP zeros give an extra ninety degrees phase delay at half the switching frequency, and an merease in gain. on (25) ~ 0 700 1k 70% ‘0 708 7k 708 00K Frequency (Ht) Fig. &: Buck Converter Current-Losp Gain. The two RHP 2t- fos a hal the switching frequency are apparent in this figure ‘The gun inereame after ths Trequency, whie the phase drops down an addtional ninety degree, If ingulficient ramp fs ade, there i very hae phase margin in this foop gu. 326C. Control-to-Output Gain A new control-ouput-voltage transfer function is Created when the current loop is closed, and the implications of the new model for current-mode control ate profound. Fig. 9 shows a plot of this transfer function for a converter with different val- ‘ves of external ramp, operating with 2 0.45 duty cycle, The aceusate representation of the system is neither first-order, as suggested in [2,3], or second order, as suggested in [J,4), but third order. Fur- thermore, the significant peaking that can oocur at half the switching frequency means that the effect of the complex poles needs to be considered even with low crossover-fequency systems, If an integral-and-lead network were added to the system for compensation, the maximum crossover fre- quency without instability would be about 3 kHz. ‘The new mode! highlights the role of the external ramp, which is used to control the Q of the second-order pole at half the switching frequency, @y- A stall extemal ramp results in high peaking of the control transfer function. When compen- sation is added to the control, this peaking will de- termine the maximum crossover frequency before subharmonic oscillations occur. ‘The approximate control-to-output transfer func- tion for the buck converter, for small values of ex- temal ramp (m_< 2), is given by EAR HORS a where S+5Ci Foes as) and 1 A, 1 (8) vor na (19) on * oe where: (20) The approximate control-output transfer function is very Useful for design purposes. Eq. (20) allows suitable choice of external ramp to prevent peaking at half the switching frequency. Design strategies ‘cant be used where a large external ramp is added 10 split the double pole, and use the low-frequency pole to cance! «zer0 of the controloutput transfer fonction. ig, 9 Contre “Ths plat 'shows the tranciton ftom sarrentmnoge sonal to ‘ohigemode conrl, ar more ermal ramp, Sade The Seubiepok! Aha the swicingtreqoney ir apparent AZ = prichng sr apparent A tore tp ade, the doubie poe bape ane verily ‘pits ito two real poles. One of ese pees then merges wit Teiow tequeney pole i form the Lecter double pote of ‘chagemode conor, andthe other mover ot bcyond Rl he ching eebeney D. Audio Susceptibitity The audio susceptibility of the buck converter shows one of the most interesting properties of current-mode control. Since the feedforward term, iy, has a negative value, it is possible to completely nll the circuit response fo input voltage portrba- tions with 4 suitable choice of extemal ramp. The ruil in audio susceptibility occurs with an external ramp value S_= S;/2. Fig. 10 shows a plot of the audio susceptibility of the buck converter as ex- temal ramp is added. The audio decreases until the all value of external ramp, then increases with the addition of further ramp. Choosing the external 327ramp 10 null the audio susceptibility can be useful for applications where output noise is extremely ritical. However, the audio susceptibility is very sensitive to changing values of the extemal ramp around this null value, and it can be difficult obtain 2 precise null. The’ approximate audio transfer function for the buck converter is 8 _ DimD ~ (1 = D/H} FOO cy Ft om= 1m * 2 Fraquoney (8) Fig 10: Andio Seep ofthe Rock Converter. This plot we te traeton from correnvmode. som to velar: ‘mose contra, 3b mor exeral ramp Stade Toe dowble Bole al the oeting regency agin apparent he Dock Sree ade ra pecan he th Poa {srbaion can stay be led by Oe sdaion of te exera ‘amp. Thee din tothe ftedorward term, hp The sudo tra ‘ey mntive foncton ef tbe ener raza’ around the mall wake 5-5/2 E, Output Impedance Closing the current feedback loop has a strong ef fect on the output impedance of the converter. With high curent feedback, the output impedance of the buck converter looks like the impedance of just the load capacitor and resistor. The significant differences from the output impedance of the ‘open-loop converter are the high de value, and the absence of any resonant peaking. Fig. 11 shows the output impedance of the buck converter with in- sreasing values of m.. Even with low levels of cur- rent feedback, there is a significant effect on the low-frequency asymptote and damping of the ‘LC-filter resonance. ‘The approximate output impedance transfer func- tion for the buck converter is 22) Z(NeRF sy Fig. 11: Output Umpedance of the Back Cawverter. This plot ‘ows the effect on output i "ie more exiernal Fam, Sys added. Carrent rode coniel giver high output impedance st low freqieney, but has no peaking atthe fter resonant fre: ‘quency. Even avery low fees of current fedback, = 32, there lea song eTect on the low-equency Output impedance VIII. Experimental Verification A converter was built with the same component values as those given for the example in the previ- us section. It was necessary to increase the input voltage 10 14 V to achieve a duty cycle of 0.45, Circuit inefficiencies and semiconductor voltage drops, not modeled in the analysis, accounted for the increased input voltage. All of the approximate ‘expressions of the previous section are functions of duty cycle, not input voltage, so the change in input voltage should not introduce any discrepancies be- tween measurements and predictions. The measured and predicted current loop gains for m_= 1 and m, = 2 are shown in Fig. 12. Both the again and phase measurements agreed very well with predictions up to half the switching frequency. It is important to point out that a digital modulator [11] was used to measure the loop gain. This ensures that the correct sampled-data loop gain is obtained [12], and that all the feedback paths created by the current loop are measured. All other measurements in this section were performed with conventional analog measurement schemes. 328° 40 2 = 0 700) 7k TOK 108K ‘0 709 7% Toe 700 Pra sl 100 2 ° 2 100 as 20 o 700 a Tor tbo Fregweccy Fig. 12: Buck Converter Curren-Loop Gain. Except for small discrepancies in the low-frequency gzin and a lower Q a the ‘Ser resonant Srequeney, the theoretical aod measured gain agree very well The mearured and predicted phate are also very dom, ‘The control-to-output loop gain, measured with the current loop closed, is shown in Fig. 13. The gain and phase measurements again show very good comrelation with the theoretical results. The peak- ing of the gain at half the switching frequency clearly shows the existence of two complex poles. This control-to-output measurement allows the voltage feedback compensation to be properly de- signed, and the onset of subharmonic oscillation can be predicted. Fig. 13: Control Output Transfer Function for Back Converter. ‘The effect of the second-order Poles is elealy shown inthe @x- _perimental measurements when no external ramp i Dyed The measurements of audio susceptibility are shown in Fig. 14. The theory and experiment agree vesy well with no extemal samp, but measurements ‘were difficult to obtain as the audio susceptibility became lower with more extemal ramp. Higher frequency measurements below -55 4B were unre- liable, dve to noise and grounding problems. How- ever, the nulling effect of the extemal ramp can be experimentally verified, with the audio susceptibility Gecreasing to a very low minimum value, then in- creasing again with more extemal ramp. The ‘measurements were extremely sensitive around the value m,= 1.5 , with small variations in the ramp causing large changes in the audio susceptibility ‘The phase of the audio measurement, not shown in this figure, switched from -180 degrees to 2210 degrees as the extemal ramp increased through its null value. This switch in polarity of the audio is predicted by the new small-signal model. 329Fig. 14: Audio Suscepiblty of the Buck Converter. The exper- {mental results show the reduction in audio ar external ramp ie inereaved from zero, then an increase af more than the nalling value is aided. Meaturements below -$5 8 were very dificult to obtain, and this is the reason for te discrepancies 2t high Frequencies with m= 1.5 0K "0K 40 700 1k Fig. 15: Outpat Impedance of the Buck Converter. Measured and predicted resuls agree wel, showing the fstorder re sponse of the output impedance with no resonant pking. The measured and predicted output impedance agreed well with external ramp adéed to give values of m, from 5 10 4. The second-order poles at half the switching frequency are not apparent in this measurement. 3X. Conchusions ‘A new current-mode control model for constant- frequency operation, which is accurate at frequen- cies from de to balf the switching frequency, has ‘been destribed Using simple pole-tero transfer functions, the model is able to predict subharmonic oscillation, avoiding the need for discrete-time 2- transform ‘models. The accuracy of sampled-data modeling is incorporated into the new model by an accurate second-order representation of the sampled-data transfer function up to half the switching frequency. Several new observations are shown about the converter systems with current-mode control. The current loop gain has a pair of complex RHP zeros which cause the instability in this loop when the extemal ramp is too small. The control-to-output transfer functions of two-state converters are best modeled by a three-pole expression. Two of these poles are at half the switching frequency, with high Q when no extemal ramp is used. The peaking at this frequency can be damped with the addition of extemal ramp, eventually splitting the poles on the real axis. ‘The new model can easily be built into any circuit analysis program such as SPICE, and can be used to. show the transition from ‘current-mode to voltage-mode control as the external samp of the Predictions of current loop output, output impedance, and audio susceptibility transfer functions are confirmed with measurements on a buck converter. It is shown that the audio susceptibility of the buck converter can be nulled with the appropriate value of extemal ramp. Acknowledgements ‘The author wishes to thank Dr. Vatche Vorperian, for many hours of vatuable discussion on current- mode control. References 1. R.D. Middlebrook, “Topics in Multipte- Loop Regulators’ and Current-Mode Programming,” IEEE Power Electronics Specialists Conference, June 24-28, 1985. 2 GC. Verphese, CA. Bruzos, KN. Mahabir, “Averaged and Sampled-Data Models for Current Mode Control: A Reexamination,” IEEE Power Electron- 330ics Specialists Conference, June 26-29, 1989, LH. Dizon, “Closing the Feedback Loop,” Appendix C, Unitrode Power Supply Desi Seminer, pp. 2C1-2C18, 19% F.C, Lee, MF. Mahmoud, ¥. Yu, “De- sign Handbook for a Standatdized Cor trol Module for DC-to-DC Converters Volume 1) NASA” CR:I6SI72, Ap J AR. Brown, R.D. Middlebrook, “Sampled-Data Modeling of Switching Regulators,” Power Electronics Special- ists Conference, pp. 349-369, June 29-Suly 3, 1981. F.C. Lee, §. Rahiaan, R.A. Carter, CH. Wu, Y. Yu, R. Chang, “Modeling and Analysis of Power Processing Systems,” Volume I - Technical Report, NASA CR-165538, December, 1980. V. Vorperian, “Simplified Analysis of PWM Converters Using the Model of the PWM Switch: Pars {and Hi,” IEEE ‘Transaction: on Aerospace and Elec- tronic Systems, March 1990, Vol. 26, No. 10. u. 12 331 2, See also VPEC Newsletter “Cunrent,” Fall 1988, and Spring 1989 Issues, Virginia Polytechnic Institute and State University. RD. Middlebrook, “Predicting Modulator Phase Lag in PWM Converter Feedback Loops,” Powercon 8, April 27-30, 1981. KJ. Astrom, B. Wittenmark, “Computer Controlled " Systems,” pp. 66-79, Prentice-Hall, Inc., 1984, R. Ridley, “Modulator Gains for Current-Mode Control”, Intemal Tech- nical Notes, VPEC, Virginia Polytechnic Institute and State’ University, 1989, to be published. BH. Cho, F.C. Lee, “Measurement of Loop Gain with the Digital Modulator,” IEEE Power Electronics Specialists Con- ference, June 18-2, 1984, AR, Brown, “Topics in the Analysis, ‘Measurement, Design of | High- Performance " Switching Regulators,” PhD. Dissertation, California Institute of Technology, May 15, 1981