Computer-Aided Design of Low Noise Microwave Circuits Thesis by Scott William Wedge In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy California Institute of Technology Pasadena, California 1991 (Submitted May 23, 1991)To Cindy, with Loveiii Acknowledgements It is a pleasure to acknowledge my advisor David Rutledge for his guidance, support, and abundant encouragement given over the years. His energy and en- thusiasm have been contagious, and have made my stay at Caltech a stimulating experience. Thanks go to Hughes Aircraft Company, in general, for being a source of continued financial support through their doctoral fellowship program. Thanks g0, in particular, to members of the RF Technology Department for tolerating my stochastic work schedule, and for listening to many wild ideas. I am grateful for the friendship, good conversation, and technical advice given by past and present members of the Caltech MMIC group. Namely, Kent Potter, Moonil Kim, Jon Hacker, Yong Guo, Zoya Popovié, Gabriel Rebeiz, Rick Compton, and Karen Lee. Extra appreciation goes to Wyman Williams for his good counsel and TgXpertise, to office mates Bobby Weikle and Victor Lubecke for often sharing their guitars, and to Phil and Tricia Stimson for too often sharing their apartment for group and bridge-night activities. New members Mike DeLisio, David Haub, and J.C. Chiao are to be thanked for the vigor they have added to the group. Additional gratitude goes to Dale Yee and Mark Vaughan for all their help on the PUFF project, and to Dale for assistance on many other matters. Finally, with unequaled gratefulness, I thank my wife Cindy and our fami- lies for tolerating my long absences, inattention, and cross country treks in the pursuit of higher education. Their love and encouragement shall not be forgotten.iv Computer-Aided Design of Low Noise Microwave Circuits Abstract Devoid of most natural and manmade noise, microwave frequencies have detection sensitivities limited by internally generated receiver noise. Low-noise amplifiers are therefore critical components in radio astronomical antennas, com- munications links, radar systems, and even home satellite dishes. A general tech- nique to accurately predict the noise performance of microwave circuits has been lacking. Current noise analysis methods have been limited to specific circuit topologies or neglect correlation, a strong effect in microwave devices. Presented here are generalized methods, developed for computer-aided design implementa- tion, for the analysis of linear noisy microwave circuits comprised of arbitrarily interconnected components. Included are descriptions of efficient algorithms for the simultaneous analysis of noisy and deterministic circuit parameters based on a wave variable approach. The methods are therefore particularly suited to microwave and millimeter-wave circuits. Noise contributions from lossy passive components and active components with electronic noise are considered. Also presented is a new technique for the measurement of device noise characteristics that offers several advantages over current measurement methods.v Contents Acknowledgements Abstract Chapter 1. Introduction 1.1 Introduction to Microwave Computer-Aided Design ................0006 4 1.2 Organization of the Thesis Chapter 2. Representation of Noisy Circuits 2.1 Scattering Parameters and Traveling Noise Waves ... 2.2 The Noise Wave Correlation Matrix 2.3 Transformations 2.4 Indefinite Multiport Representations 2.5 Analysis of Noisy Two-Ports Chapter 3. Linear Connection Theory ...............:.0e0seeeeecees 32 3.1 Multiport Wave Variable Connection Methods ...............000000008 33 3.2 Network Reduction by Subnetwork Growth ............606e0ceeeeeeeee 38, 3.3 Embedded Network Connections Chapter 4. Component Modeling 4.1 Noise Waves and Passive Components ... 4.2 Passive Component Parameters 4.2.1 Dispersion 4.2.2 Losses .. 4.2.3 Quality Factor .. 4.3 Active Device Modeling . Chapter 5. Noise Wave Measurement . 5.1 Tuner Extraction of Noise Parameters ... 5.2 One-Port Noise Wave Measurements 5.3 Two-Port Noise Wave Measurement ... Chapter 6. Conclusions and Suggestions for Future Work ......... 98Chapter 1 Introduction Microwave frequencies occupy a portion of the electromagnetic spectrum that possesses an unusual quality. Lower frequencies are plagued by urban, galactic, solar and atmospheric noise, while higher frequencies suffer from at- mospheric attenuation. In the microwave band, falling between 300 MHz and 30 GHz, nearly every type of natural and manmade noise is absent. This low- noise nature permits the detection of very faint signals; a tremendous advantage for the wireless transmission and/or reception of energy. Exploitation of this property has led to successful terrestrial, satellite and deep-space communica- tions, ultra-sensitive radio astronomy, and accurate radar detection, tracking, and guidance. In order to take full advantage of the noise absence at microwave frequencies, receivers require front-end amplifiers that contribute little of their own noise. In recent years, a considerable amount of research has been pursued specifically toward the development of microwave and millimeter-wave transistors with im- proved noise performance [1]. A summary of the current state-of-the-art is given in Fig. 1.1, where noise temperature for microwave transistors is plotted with respect to frequency. A room temperature bipolar transistor is shown in compar- ison to room temperature and cryogenic gallium arsenide field effect transistors (GaAs FET’s) and high electron mobility transistors (HEMT’s). Also included in the figure are sources of external noise in and about the microwave band that2 place a fundamental limit on receiver noise temperature. At low frequencies, galactic noise dominates. At high frequencies, the atmospheric absorption ef- fects are evident, due primarily to oxygen and water vapor. In much of the microwave band the only limitation to low noise performance is a small amount of atmospheric absorption and the 3 K cosmic microwave background radiation: a remnant of the intense radiation associated with the Big Bang. The cryogenic HEMT stands out as a superior device for low-noise amplification. It provides an alternative to 4 K masers at an enormous reduction in cost [2]. Refrigeration is impractical in most microwave systems, so room temperature HEMT operation near the background radiation limit is the objective of ongoing research. Recent years have also witnessed dramatic increases in the development and use of microwave computer-aided design (CAD) tools. This has had a profound effect on microwave circuit design strategies. Lab bench tweaking in both the de- velopment phase and on the production line have become forbidden in the pursuit of cost effective engineering. Experimentation and development now take place using computer simulations. Efficient algorithms and more accurate component models are continually sought to ensure fast and precise performance prediction. The advent of monolithic microwave integrated circuits has also hastened the development of CAD tools. These allow little on-chip testing and cannot be tweaked or repaired to correct poor performance. Accurate CAD is needed to insure successful first-time chip designs. While tremendous accomplishments have been made in low-noise transistor technology and microwave CAD development, noise analysis has long been a miss- ing aspect in computer-aided design. When it has been offered, it has typically been restricted to circuits with specific topologies [3], or the effects of correlated noise have been neglected [4]. As a result, the noise performance of many ampli- fiers could not be accurately predicted prior to fabrication. Only recently have noise analysis methods emerged with the ability to analyze microwave circuits1000 Ey oot Galactic Noise 300K Bipolar 100 300K FET Atmosphere (5K FET Noise Temperature (K) 3K Cosmic 4 |. Background 0.1 1 10 100 Frequency (GHz) Figure 1.1 Noise temperatures for state-of-the-art microwave transistors compared to natural noise limits.4 comprised of arbitrarily interconnected components [5]. Presented here is a new wave variable approach. It is particularly suited to microwave and millimeter- wave circuits which are characterized in terms of wave variables. An integral part of the noise analysis process is the characterization of active microwave devices by both modeling and measurement [6]. New methods to accomplish this us- ing wave variables are also presented. These methods are shown to offer several advantages over current techniques. 1.1 Introduction to Microwave Computer-Aided Design Predicting the performance of a low noise microwave circuit involves a linear frequency domain analysis. The basic problem is illustrated by Fig. 1.2: given a collection of components, depicted in the figure as S1, $2, ...; Sm, the objective is to determine the response of the aggregate network, shown as Snet, based on knowledge of the circuit’s topology (i.e. the interconnection of the components), and its excitations. The typical computer algorithm obtains a solution by describ- ing the network in terms of Kirchhoff’s voltage and current laws and component constitutive relations. This description is used to form one large, although often sparse, matrix equation which when solved gives all values for voltage and cur- rent at every node and branch. The large matrix equation is called the tableau. Sparse matrix means for its solution are referred to as sparse-tableau methods [7]. A microwave circuit has physical dimensions comparable to a wavelength. Under such conditions voltage and current behave as traveling wave quantities, with impedance mismatches resulting in reflections and standing waves. Voltage and current variations within a circuit can therefore be quite dramatic. Con- sequently, sparse-tableau methods using voltage and current variables are often numerically unstable. A preferable procedure uses traveling wave variables [8]. In the wave approach, scattering parameters are derived from complex ratios of the wave variables. These parameters form scattering matrices whose manipula- tion is computationally stable. In the simple example of a one-port element withFigure 1.2 A collection of components, 51, 52, ..., Sm, interconnected to form network Suet. Solid lines depict external connections; dotted lines, internal connections. impedance Z(w) terminating a transmission line with normalizing impedance Zo, the scattering parameter $(w) is given by Zw) — 25 )= Za) F a" a4) If the impedance Z(w) is from a passive component, its value will lie in the half- plane given by Re(Z(w)) > 0. The bilinear transform (1.1) maps this half-plane into the unit circle |$(w)| < 1. If Z(w) was the impedance of a parallel LC circuit, it would pass through a resonance where |Z(w)| tended to infinity. This would S(w result in considerable numerical difficulties if an impedance representation were used. Mapping into the unit circle causes scattering parameters, in comparison, to be well behaved. Since resonances are common in high frequency circuits, the improved numerical stability is welcome.6 Scattering matrices have the additional advantage of existing for nearly all networks. Exceptions are caused only by the presence of specific values of nega- tive resistance. In voltage-current methods, nonexistence is a frequent problem and has led to the increased complexity of the modified nodal approach [9]. Sim- plification is possible using graph theoretic derivation of the tableau [10]. Yet, even simpler graphical methods are possible in the wave approach using Ma- son’s theory of signal-flow graphs [11]. It shall be shown that Mason’s theory leads to powerful methods for network reduction and subsequent calculation. ‘These advantages, combined with the fact that scattering parameters are ob- tained directly during the measurement process, have made the wave approach the preferred method for both microwave component characterization and net- work analysis (12]. The superiority of the wave approach has led to the hypothesis that a compa- rable means for performing noise analysis would share its benefits. Noise analysis using wave variables has appeared only recently 13,14]. Much improvement in these methods is possible through application of the connection theory presented here. Connection theory has been shown to result in considerable reductions in computer time and memory allocation needs [8]. Historically, it may be traced to Murray-Lasso [15] who conceived the idea of dividing admittance networks into smaller subcircuits to simplify the calculation of the aggregate network. The promise of this idea may be understood by a second look at Fig, 1.2. The connections internal to the network Snet are shown with dotted lines. Those con- nections external to the network, accessible by terminals, are shown in solid lines. Connection theory obtains simplification by avoiding storage and direct calcula- tion of variables associated with the internal connections. Internal and external variables are separated to form a tableau whose solution is derived from a lower order matrix inversion. Such a method using wave variables is given here. It7 allows the efficient and simultaneous calculation of both deterministic and noise parameters for interconnected microwave networks of arbitrary topology. Some of the research described here is based on the author’s contributions to the interactive microwave circuit analysis package PUFF [16]. Since 1987, this software and accompanying manual have been distributed to over 8000 users in 40 countries. PUFF is used to synthesize, analyze, and lay out microstrip and stripline circuits. It is limited, however, to the analysis of deterministic circuits, and cannot perform noise analysis. The features of PUFF, plus a noise wave analysis capability, have been combined to form another circuit simulator named DRACO. The programs are quite similar, the primary difference being DRACO's ability to perform noise analysis. 1.2 Organization of the Thesis Given in Fig. 1.3 is a simplified, generic, computer-aided analysis algorithm for the calculation of signal and noise parameters for a network. The two major steps involved are calculation of noise and deterministic parameters for each com- ponent, and the application of connection theory to solve for the parameters of the aggregate network. Detailed descriptions of how these steps are accomplished are the subject of this work. In chapter 2, wave representations for signals and noise are discussed. De- fined here is the noise wave correlation matrix used to characterize the statistics of noise sources. Basic manipulations of correlation matrices are described, as well as relationships to two-port figures of merit. Chapter 3 presents linear connection theory applied to both signal and noise calculations using wave representations. A detailed description of the subnetwork growth method is given: the process wherein calculation of a network is made by repeated connection calculations. Chapter 4 is concerned with the derivation of signal and noise parameters for the basic components used in microwave circuits. These include single and8 coupled transmission lines, and microwave transistors. It is shown that in many cases the noise wave parameters can be derived directly from scattering param- eters. The current methods used for the measurement of noise parameters for mi- crowave devices are known to have several limitations (6]. Presented in chapter 5 is an alternative technique used to make direct measurement of noise wave pa- rameters. Proposed in chapter 6 are research directions in noise analysis and measure- ment that should be addressed. These areas concern the modeling of microwave devices for noise, and new measurement approaches. Start | Calculate noise | and deterministic parameters for each component { Increment frequency Apply connection theory to compute parameters for | aggregate network Stop Figure 1.3 Generic computer-aided analysis algorithm for linear frequency domain anal- ysis.[10] uy 9 References S. Vaughn, K. White, U.K. Mishra, M.J. Delaney, P. Greiling, and S. Rosen- baum, “High-performance V-band low noise amplifiers,” IEEE MTT-S Symp. Dig., pp. 801-804, 1989. M.W. Pospieszalski, S. Weinreb, R.D. Norrod, and R. Harris, “FET’s and HEMT’s at cryogenic temperatures — their properties and use in low-noise amplifiers,” IEEE Trans. Microwave Theory Tech., vol. MTT-36, pp. 552-560, Mar. 1988. H. Hillbrand and P.H. Russer, “An efficient method for computer aided noise analysis of linear amplifier networks,” IEEE Trans. Circuits Syst., vol. CAS-23, pp. 235-238, April 1976. R. Rohrer, L. Nagel, R. Meyer, and L. Weber, “Computationally efficient electronic-circuit noise calculations,” IEEE J. Solid-State Circuits, vol. SC-6, pp. 204-213, Aug. 1971. V. Rizzoli and A. Lipparini, “Computer-aided noise analysis of linear mul- tiport networks of arbitrary topology,” IEEE Trans. Microwave Theory Tech., vol. MTT-33, pp. 1507-1512, Dec. 1985. A. Cappy, “Noise modeling and measurement techniques,” IEEE Trans. Mi- crowave Theory Tech., vol. MTT-36, pp. 1-10, Jan. 1988. G. D. Hachtel, R. K. Brayton, and F. G. Gustavson, “The sparse tableau approach to network analysis and design,” IEEE Trans. Circuit Theory, vol. CT-18, pp. 101-113, Jan. 1971. V. A. Monaco and P. Tiberio, “Computer-aided analysis of microwave cir- cuits,” IEEE Trans. Microwave Theory Tech., vol. MTT-22, pp. 249-263, Mar. 1974. C. Ho, A.E. Ruehli, and P.A. Brennan, “The modified nodal approach to network analysis,” IEEE Trans. Circuits Syst., vol. CAS-25, pp. 504-509, June 1975. J. Viach and K. Singhal, Computer Methods for Circuit Analysis and Design, Van Nostrand Reinhold, New York, 1983. S.J. Mason, “Feedback theory — further properties of signal flow graphs,” Proc. IRE, vol. 44, pp. 920-926, July 1956.{22} 23] {14} (15) (16) 10 K.C. Gupta, R. Garg, and R. Chadha, Computer-Aided Design of Microwave Circuits, Artech House, Norwood, Massachusetts, 1981. N.G. Kanaglekar, R.E. McIntosh, and W.E. Bryant, “Wave analysis of noise in interconnected multiport networks,” IEEE Trans. Microwave Theory Tech., vol. MTT-35, pp. 112-115, Feb. 1987. J.A. Dobrowolski, “A CAD-oriented method for noise figure computation of two-ports with any internal topology,” IEEE Trans. Microwave Theory Tech., vol. MTT-37, pp. 15-20, Jan. 1989. M.A. Murray-Lasso, “Black-box models for linear integrated circuits,” IEEE Trans. Educ., vol. E-12, pp. 170-180, Sept. 1969. R. Compton, 8.W. Wedge, and D. Rutledge, PUFF: Computer-Aided Design for Microwave Integrated Circuits, published at Caltech, Pasadena, California, Jan, 1990.il Chapter 2 Representation of Noisy Circuits A large body of work has been devoted to noise representations for two-port circuits. Hartmann [1] surveys twelve different methods, showing that each re- quires four noise and four complex signal parameters per frequency. The use of multiple representations is common because each offers a simple noise calculation for components connected in a specific manner. The computer-aided noise anal- ysis strategy has been to use the appropriate noise representation for the type of connection to be made [2]. The advantages are akin to deterministic methods: admittance matrices allow a simple calculation for parallel circuits, impedance matrices simplify the series connection, and the chain representation simplifies the cascade connection. This approach works well for interconnected two-ports. For arbitrarily interconnected multiports, however, the types of connections and hence the number of definable representations are essentially unlimited. For the analysis of arbitrarily complex circuits, the representation must be applicable to any number of ports. The chain representation, for example, uses both noise voltage and current sources and has a multiport equivalent only when an even number of ports is considered [3]. The deterministic and noise proper- ties of a circuit must be characterized in a compatible manner. Noise modeling with an admittance matrix, for example, requires noise current sources, while noise voltage sources would be used with an impedance matrix. As discussed in chapter 1, these representations use ratios of voltage and current that can be12 numerically unstable in microwave circuits. Scattering matrices are considered superior. In contrast to the traditional use of noise voltage and current sources, a representation compatible with scattering parameters shall be developed here using noise wave sources. Wave representations for noise are not new. Bauer and Rothe [4] and Penfield [5] gave examples soon after Rothe and Dahlke pub- lished their definitive theory of noisy fourpoles (6] using correlated voltage and current sources. Bosma [7], using a noise wave approach, has reworked the Haus and Adler [3] optimum noise performance problem. The noise wave definitions given here are similar to those used by Bosma. They allow generalized analysis methods, and parameterization based on statistical properties. 2.1 Scattering Parameters and Traveling Noise Waves Given in Fig. 2.1 is a transmission line of characteristic impedance Zo ter- minated in an impedance Z(w). A deterministic voltage wave V+, traveling in the positive z direction, is incident from the left. It is scattered, resulting in the voltage wave V.. traveling in the opposite direction. In phasor notation, the deterministic voltage waves Vj. and V_ have the form Vy. = Acitt e58= (2.14) V_ = Beit- ei?* (2.18) where A and B are rms magnitudes, f is the propagation constant, and $4 and ~ are phases. The scattering parameter representation is made in reference to normalized values of these voltage waves, given by Vey 0 Veo where Zo is the real characteristic or normalizing impedance of the line, a is the (2.2) incident wave, and b the scattered wave. This normalization results in the power propagating in the positive and negative z direction to be given by |a|? and |6[?,13 respectively. The scattering parameter S is the ratio of the two traveling waves, and the one-port circuit of Fig 2.1 is represented by the simple expression b= Sa (2.3) where S may be found from (1.1). Figure 2.1 A transmission line with characteristic impedance Zo terminated in an impedance Z(w). Voltage wave Vj, is incident, causing scattering wave V_. Wave Vq— is due to noise. The above result does not account for noise that could be produced by the terminating impedance. For example, a passive Z(w) would produce an output wave due to its thermal noise. This wave will be a random variable, but if confined to a small bandwidth Af about a mean frequency f, it may be written with a phasor defined at @ = 2rf in a manner similar to the deterministic waves, namely Va—(t) = C(t)eien(®) eB (2.4) where B is the mean propagation constant at f. Random variables C(t) and én(t) describe the rms amplitude and phase, respectively, of the wave envelope. Assuming Af < f, these will be slowly varying functions. This leads to the definition of a normalized noise wave c given by (2.5)14 Since V,_(t) is given in terms of an rms phasor, |c|? will have units of power, as in the case of |a|? and |8|?. Although c is a time varying complex random variable, its randomness and time dependence are removed by its characterization in terms of time averaged power |c|? given by [8] apt = inal, le(s)[2ae. (26) In general, the value of |e[? will be a function of both f and Af. Based on these definitions, it is now possible to rewrite (2.3) to include the effects of the one-port termination’s noise. This is done by adding a wave source term c: b=Sate. (2.7) ‘The output wave 6 now includes contributions from both the scattered wave Sa, and the noise wave. A complete representation for the one-port now includes the scattering parameter S and the noise power |c|? that is available from the termination. ‘The result of (2.7) can be generalized to the case of an arbitrary multi- port network, Given in Fig. 2.2 is a multiport network S with incident waves @1,42,...,4, and output waves bj, b2,...,b, present at each terminal. As in the case of the one port, these are normalized traveling waves defined as in (2.2). Elements of the scattering matrix, or S-matrix, are defined in terms of complex ratios of these waves (2.8) where noise sources are neglected and the restriction that a; is the only incident wave implies that all other ports are terminated in the normalizing impedance Zo. The diagonal elements, s;;, of the S-matrix thus represent reflection coefficients, while the off-diagonal elements, s;;, represent either gain or isolation.15, Figure 2.2 A multiport network S with incident waves a1,2,...,@q, output waves bi, bo, ..-4bns and noise waves ¢1,¢2,.-.,Cn, Present at each port. To model the effects of noise, Fig. 2.2 includes wave sources c1,¢2,..-;Cn that represent contributions to the output waves due to the multiport’s internal noise. These noise waves are outwardly directed correlated sources present at each port, defined as in (2.5), and represent the noise originating in the network that is deliverable to its terminations. For the general multiport element, the relation between incident and output waves is therefore given’ by the matrix equation b=Sate (2.9) where a and b are vectors denoting incident and output waves, respectively, S is the scattering matrix, and the noise waves are given by vector c. A physical interpretation of the noise waves is in order. Based on the definition of (2.9) they may be found mathematically by satisfying c=b] . (2.10) Jao Itis therefore tempting to define c as those waves the network S delivers into non- reflective terminations in the absence of input. Yet, physically such an absence is16 unrealizable: the multiport would have to be terminated in non-reflective loads (ie. characteristic impedances Zp) that produce no noise, implying resistive terminations at 0 K. Noise emanating from the terminations will always yield a nonzero a vector which in general will be scattered by S and contribute to b. The noise wave vector ¢ is therefore best interpreted as the excess noise wave that the multiport delivers to its terminations. 2.2 The Noise Wave Correlation Matrix As in the case of the one-port, the components of ¢ are complex random variables. Parameterization is required in terms of their statistical properties. This is accomplished through definition of the noise wave correlation matrix C, given by (2.11) where the dagger indicates Hermitian conjugation, and the overbar indicates statistical averaging. The components of the matrix are the correlation products cic 4 5 faa | cx(t)eF(t)at (2.12) where, as before, 0767 is in general a function of both f and Af. For the arbitrary multiport of Fig. 2.2, the noise wave correlation matrix is written lex? aac cy C2ch (2.13) eel Gach. [end For each noise-wave ¢; the correlation matrix yields values for the noise power deliverable to the terminations, given by Je//?, as well as a correlation product with each of the other ports, given by ic}. An n-port linear component’s signal and noise properties are completely described by n xn scattering and noise wave correlation matrices.17 Specific conditions must be met in order for the noise wave correlation matrix to exist. Noise waves are defined in a bandwidth Af about a mean frequency Ff, so a frequency domain representation of the statistics of the waves must be possible. This is guaranteed by considering only noise processes that are wide- sense stationary. When more than one process is considered (i.e. cross-correlation between noise waves) these must be jointly wide-sense stationary. The processes must also be ergodic to ensure that the time limits given in (2.6) and (2.12) exist and correspond to ensemble averages. When two noise processes c; and c; meet these conditions, the correlation product ¢j¢} can be represented approximately by FS ASWy(F) (2.14) where Wi; is the self (i = j) or cross (i # j) power spectral density of the two processes, assumed to have little variation over the bandwidth Af. The factor of two accounts for the two-sidedness of the spectral density function. The approximation may be made arbitrarily close by letting Af approach zero. For the purpose of numerical manipulation the Af is often dropped. An assumption is made either that Af = 1 Hz or that the correlation terms are power density. A power density interpretation changes approximation (2.14) to an exact relation. In either case, the values of the correlation matrix are referred to as spot noise values. Values of Af can be significant when related to measurement of the noise waves. Coherence length and time will be set by the measurement system. A system with group velocity v, will have a coherence length given by A€ = v,/Af and coherence time At = 1/Af. The dimensions and bandwidth of the mea- surement system must be such that the random variables of (2.4) are sufficiently slowly varying to allow noise wave comparisons. A transmission line system will typically have multiple reflections that must be taken into account. Requiring18 phase changes due to coherence length and time to be much less than 27 results in the criteria nL < Ae (2.15) where n is an integer accounting for the number of reflections, and L is the phys- ical dimension of the measurement system that accounts for wave delays. A typ- ical spectral density measurement of a microwave circuit will have Af + 4 MHz, and easily satisfy (2.15). This is often combined with a long integration time to improve the estimation of (2.12). 2.3 Transformations A noise wave correlation matrix will undergo a transformation upon oper- ation by a linear system. For example, a noise wave vector c, may be modified by a transformation T resulting in a vector c' = Tc. The original noise wave correlation matrix C, = cet will therefore undergo an Hermitian congruence transformation to a new matrix given by Cc TC, (2.16) The general form of (2.16) is apparent in any transformation of a correlation matrix, Such transformations occur when components are added to a network, and shall be considered in detail in the next chapter. Changing from one noise representation to another is the equivalent of op- eration by a linear system. Such a change may be made for computational advantage or to avoid singularities. In addition to the wave approach, admit- tance and impedance representations have applications. All three are compared in Fig. 2.3 including their respective circuit diagrams and matrix relations for voltage, current, and wave quantities. Methods using impedance matrices and noise voltage sources have been described by Haus and Adler [9], while noise19 analysis using admittance matrices and noise current sources has been discussed by Rizzoli and Lipparini [10]. The correlation matrices are defined similarly Cy = inih (2.174) Cc, =vavi (2.176) where in is the noise current vector of the admittance representation and Va is the noise voltage vector of the impedance representation. Relations between wave, voltage and current quantities for both signals and noise are necessary to derive transformations. The signal traveling waves used to define the scattering matrix are related to the terminal voltage and current of the multiport by (2.184) (2.186) where normalization to the characteristic impedance has been assumed. These identities allow derivation of the signal matrix transformations given in Fig. 2.3. Comparison of noise quantities is assisted by defining the following waves in terms of the multiport’s short circuit noise current and open circuit noise voltage an = iva tin) (2.194) ba = H(vn in) (2.198) where again normalization to the characteristic impedance is assumed. The noise wave vector is then given by ¢ = bp —Saq (2.20) where S is the scattering matrix. Relations (2.19)-(2.20) lead to the family of noise correlation matrix transformations, also listed in Fig. 2.3.20 ‘srojouresed oava pure ‘ouwpodunt ‘soue;yrumpe jo uostredwioy g'z amt (S-D'0(s-Di="o (s+Dso(s + DF = "9 (Z+D*0(Z+D="9 1240Z="0 suoryeULIO}sUE IT, xEEWy ‘WOTyEpaLIO (A-Di(A+D=8 (S-I(S+D=Z (St D(S-D=A suOreUrIO}streyy, xu (I-2Z);-1+Z) =s rA=Z r-Z=K puss #2="0 gata =" i= sonnuopy yexous5) a+es=q MAFIZ=A THAXR=I > q ureaSerq, +e - yong o aARy souepedury souryrupy | woreyuasardoy21 2.4 Indefinite Multiport Representations Signal matrix and noise correlation matrix data derived from swept fre- quency measurements are often used in the computer-aided design process. Of ten these data are unsuitable for direct numerical manipulation. Active devices are generally measured with one terminal grounded, and the resulting definite parameters, so named because a clearly defined ground connection exists, char- acterize the N terminal device in terms of an N —1 multiport. In the case of transistors, two-port common source or common emitter measurements charac- terize the three terminal device. These definite parameters must be converted to a form that allows the device to be placed in any configuration, and not simply the common source or common emitter configuration in which it was measured. Indefinite parameters provide the solution. The indefinite parameters of an N terminal device may be derived from its definite parameters by mathematically changing the grounded terminal into an accessible terminal. This yields the desired N-port characterization of the N terminal device. An indefinite multiport is defined as one in which no internal connection to ground exists. The terminal currents then satisfy Kirchhoff’s current law (KCL) N Yi; =0. (2.21) At microwave frequencies this condition is not assured. Currents to ground resulting from stray capacitance would not pass through a terminal, and would result in errors. When KCL is valid, the use of an impedance representation is precluded; indefinite impedance matrices do not exist. In terms of admittance and scattering parameters, KCL results in useful properties. For an indefinite admittance matrix, (2.21) causes the sum of any row or column in the matrix to equal zero, the components satisfying N. N > Uik = > yjk = 0. (2.22) ja k =122 ‘This result facilitates the conversion from definite to indefinite parameters. Con- necting a terminal of a multiport admittance to ground will reduce its matrix by one row and one column, but will not affect the remaining components. The definite and indefinite matrix components y;x will be identical for j,k < N. The definite parameters fill most of the indefinite matrix; only a single row and column need be generated, and this is possible using (2.22). In the case of scattering parameters, the enforcement of KCL results in the sum of any row or column in the indefinite scattering matrix to equal unity; the components satisfying (2.23) Ny a eee 5 im Since definite and indefinite scattering matrices generally do not share common components, (2.23) does not allow a simple conversion between the two. Instead it permits N — 1 port scattering parameter measurements to be used to fill an N port indefinite matrix if the N’th port of the device is terminated in the normalizing impedance during measurement [11]. Parameters can then be shared, and the N’th row and column of the indefinite matrix would be filled using (2.23). When measurements are performed with the N’'th terminal connected to ground, however, the conversion to an indefinite scattering matrix is best made using transformations to and from the admittance representation given in Fig. 2.3. " Multiport noise correlation matrices may also be converted from definite to indefinite forms. Based on the general identities given in Fig. 2.3, the noise currents in an indefinite multiport admittance must also satisfy KCL. It then fol lows that the sum of any row or column in an indefinite noise current correlation matrix Cy is equal to zero: N Viste fe (2.24)23 As in the case of the deterministic admittance matrix (for j,k < N) the values injta, are the same for definite and indefinite matrices. The indefinite noise current correlation matrix is therefore easily filled using (2.24) and the values from the definite matrix. In the case of the indefinite noise wave correlation matrix, KCL and the relations in Fig. 2.3 lead to the result that yoo Ladg= at an (2.25) The sum of any row or column in C, equals zero, as well. As with scattering parameters, conversion from a definite to an indefinite noise wave correlation ma- trix is difficult since they do not share common components. As before, however, if N—1 port noise measurements are performed with the N’th port terminated in the appropriate normalizing impedance, these measurements and (2.25) can be used to fill the indefinite matrix. When the noise measurements are performed with the N’th terminal grounded, the generation of the indefinite noise wave matrix is best made using the transformations given in Fig, 2.3. 2.5 Analysis of Noisy Two-Ports Although a microwave circuit may consist of a complicated arrangement of various components, the result is often a two-port device. A noise performance appraisal for the two-port is made using well known figures of merit. These may be calculated using values of the scattering and noise wave correlation matrices. Noise figure (F') has been defined by Friis [12] as a measure of the degradation in signal-to-noise ratio. It may be written as Na Falta, (2.26) where N, is the noise power added by the two-port that is available at its output, Gz is the available power gain, and N, is the thermal noise power available from a24 source assumed to be at temperature Ty = 290 K. This definition can be applied to the two-port amplifier represented by scattering parameters and noise waves as given in Fig. 2.4(a). An input signal source is shown as reflection coefficient T’,. The two-port has outward directed noise waves cy and cz. The noise waves are shown as source terms in the signal-flow graph of Fig. 2.4(b) [13]. The nodes on the graph denote traveling wave variables while the branches represent transfer (ie. scattering) coefficients. Mason’s theory [14] is used to solve for the needed quantities, The available gain of the two-port is G, = — ba? G= IPP) © fl =Tasu PC = IstaF) where s4p is the equivalent reflection coefficient seen at the output of the two-port (2.27) given by si2T',S21 TT’ (2.28) S22 = $22 + The thermal noise power available from the source is known from Nyquist’s formula to be N, = kT (2.29) where k is Boltzmann’s constant and a 1-Hz bandwidth is assumed. The available noise power added by the two-port is also found by evaluation of the signal-flow graph. The noise power at the output is found by adding contributions to node b due to c, and cz. This gives = shal (230) Substitution of the above expressions into (2.26) results in 1-Tysuy|* ot. ta Een)| F=l+ a (2.31) To(1 — |P)?)25, (b) Figure 2.4 Schematic (a) and signal-flow graph (b) diagrams for a noisy two-port com- ponent.26 ‘The numerator can be expanded and written in terms of the two-port’s noise wave correlation matrix C,. Doing so and writing the result in terms of noise temperature T, = To(F — 1) gives BC, Bt kT, = OO 2.32 a=1P) ve where f is the 1 x 2 row matrix given by ga(r, Lhe). (2.33) $21 Haus and Adler [3] have defined the figure of merit known as noise measure which determines the best cascade order of two amplifiers with respect to the noise performance. It takes into account the amplifier gain that, if small, would cause poor cascaded noise figure. Applying the definition given by Fukui (15] gives - BC.Bt = - ERE ssh (2.34) where S is the scattering matrix of the two-port, and again a 1-Hz bandwidth has been assumed. Equations (2.32) and (2.34) demonstrate that if the noise correlation ma- trix of the two-port is known, then its noise figure, noise temperature, and noise measure can be predicted as a function of the source reflection coefficient. It is not typical, however, to make direct measurement of the components of Cs. Instead, two-port measurements are commonly made to extract the noise pa- rameters known as Tmin (minimum noise temperature), I'op: (optimum source reflection coefficient), and Rp, (noise resistance). These parameters satisfy the noise temperature relation Dy = Tryin ¢ AR a Pope Z(t Foye = Fal (238)27 By comparing (2.35) and (2.32), conversion formulas between the above noise parameters and elements of the noise wave correlation matrix may be derived. These are — ket|1 — $13 Dopel? ? = kTmin 2 —1) + tert 2.36. Terk (lu -1) +" (2.364) kt|D opel? ? er optl lea? = |sail (i mnin + i +1) (2.360) = W847 Spe ht = (_KtDope? ae = Epp ten say +The t &Tmin) (2.36¢) where pt = SETORn (2.37) Zo The inverse relations are Ltsny{? kt =| —c CE) (2.38a) 2 1, = GE = sn = oP al? CED {sai |(1 + [Popel?) Pope = a(: ie a ) (2.88) iF a _ bP +a aenF (oa) TeaTsia ~ e1eoa1 where These relations are rather complicated, but some insight can be gained by defin- ing the denominator of n as a parameter ¢ given by ¢ = lea Psi — ciczsai- (2.40)28 This parameter represents a measure of the correlation between c, and cp refer- enced at an input node. The following relations exist between Pop: and ¢: fim, Penel (2.41a) arg(Dopt) = — arg(¢) (2.416) i.e., when the correlation term ¢ vanishes, Tmin has been obtained, and a 180° phase difference exists between ¢ and I'gp¢. The search for an optimum noise match using the noise wave correlation matrix thus involves finding a network that gives ¢ = 0. Since impedance matching is also of concern, this can be interpreted as a search to make cc, = 0 and 31; = 0. ‘The complexity of the conversion formulas given above is due to differences in the nature of the noise parameter definitions, each being a function of [,. As apparent in (2.35), the noise parameters Foyt, Tmins and Ry, are defined in such a way as to lead to a simple expression for T, when I', = Topt. The noise wave correlation matrix parameters, however, lead to a simple expression for I, = 0. Substitution of I, = 0 into (2.32) and (2.34) results in a Jel iT, = TOE : 2.42) i821 ( ) for noise temperature and lea)? — 2.43) Tal + eI 243) for noise measure. The single output noise power |c2|? of the two-port is the sole indicator of noise performance with a non-reflective (Zo) input termination. Using noise parameters referenced to Zo terminations permits a new low noise design process. The traditional procedure is to create a matching network that transforms a T, = 0 load to I, = Tope. This network is then connected to the input of the device to realize Tmin- This is an inside-out approach: the29 matching network and the device to be matched have been divided and analyzed separately. Noise waves permit an outside-in approach. The complete circuit, including matching network and device, is built up and analyzed as a whole. Each change will, in general, cause a change in the correlation matrix, and lead to new values for T, (2.42). The Trin design is known to be realized when ¢ = 0, or alternatively, when ¢,¢3 and s, are nulled. The approach is compatible with deterministic amplifier design using scattering parameters, where a typical ob- jective is to null s1; and sz for good input and output matching. In chapter 5, methods shall be shown for the direct measurement of noise waves in a Zp sys- tem. In general, the outside-in approach has the advantage that the parameters calculated are also those that are measured.a {10 (11) (12 (13 30 References K. Hartmann, “Noise characterization of linear circuits,” IEEE Trans. Circuit Syst., vol. CAS-23, pp. 581-590, Oct. 1976. H. Hillbrand and P.H. Russer, “An efficient method for computer aided noise analysis of linear amplifier networks,” IEEE Trans. Circuits and Systems, vol. CAS-23, pp. 235-238, April 1976. HLA. Haus and R.B. Adler, “Optimum noise performance of linear ampli- fiers,” Proc. IRE, vol. 46, pp. 1517-1533, Aug. 1958. H. Bauer and H. Rothe, “Der aquivalente Rauschvierpol als Wellenvierpol,” Arch. elekt. Ubertragung, vol. 10, pp. 241-252, June 1956. P. Penfield, “Wave representation of amplifier noise,” IRE Trans. Circuit Theory, vol. CT-9, pp. 84-86, Mar. 1962. H. Rothe and W. Dahlke, “Theory of noisy fourpoles,” Proc. IRE, vol. 44, pp. 811-818, June 1956. H. Bosma, “On the theory of linear noisy systems,” Phillips Res. Repts. Suppl., no. 10, 1967. W.B. Davenport and W.L. Root, An Introduction to the Theory of Random Signals and Noise, McGraw Hill, New York, 1958. HLA. Haus and R.B. Adler, Circuit Theory of Linear Noisy Networks, John Wiley & Sons, New York, 1959. V. Rizzoli and A. Lipparini, “Computer-aided noise analysis of linear mul- tiport networks of arbitrary topology,” IEEE Trans. Microwave Theory Tech., vol. MTT-33, pp. 1507-1512, Dec. 1985. A.P. Anderson, “High-frequency transistor evaluation by three-port scatter- ing parameters,” IEEE Trans. Microwave Theory Tech., vol. MTT-15, pp. 263- 265, April 1967. H.T. Friis, “Noise figures of radio receivers,” Proc. IRE, vol. 32, pp. 419-422, July 1944. RP. Hecken, “Analysis of linear noisy two-ports using scattering waves,” IEEE Trans. on Microwave Theory and Tech., vol. MTT-29, pp. 997-1004, Oct 1981.31 [14] S.J. Mason, “Feedback theory ~ further properties of signal flow graphs,” Proc. IRE, vol. 44, pp. 920-926, July 1956. [15] H. Fukui, “Available power gain, noise figure, and noise measure of two-ports and their graphical representation,” IEEE Trans. Circuit Theory, vol. CT-13, pp. 137-142, June 1966.32 Chapter 3 Linear Connection Theory In the wave approach to network analysis, direct application of Kirchhof's laws is inappropriate. Alternatives are sought that allow sufficient mathemat- ical circuit descriptions to permit network solution. These come in two forms: wave equivalence based on network connections, and signal-flow graph theory. Combined, these form the basis of linear connection theory. As with any approach, a network is solved based on knowledge of its com- ponents and its topology. Both brute force and elegant means of solution are typically possible. For the computer, the brute force method involves placing all network information in a large matrix called the tableau. Solution is then accom- plished by matrix inversion. In this chapter, the brute force method shall first be examined. This is followed by pursuit of more elegant means to simplify the matrix inversion problem to reduce computer time and memory space require- ments. This leads to the subnetwork growth method: a numerical application of signal-flow graph theory.33 3.1 Multiport Wave Variable Connection Methods The basic network analysis problem is illustrated by Fig. 3.1(a). A group of components 51, S2,..., Sm, have been combined to form network Spee. In the wave variable interpretation, these components are represented by scattering matrices. Associated with each multiport scattering matrix S, shall be an input wave vector a,, an output wave vector by, and an output wave source vector C,. All of these variables may be placed into a tableau equation given by by S 0... 0 ay C1 bz o&. oO a 2 eyate Po TPP yells]. ep bm 0 0 ... Sm am cm This may be simply rewritten in the form of a standard wave variable equation b=Sat+e (3.2) keeping in mind that a, b, and c, are now composite wave vectors, and S is a block diagonal matrix whose submatrices along the diagonal are the scattering matrices of the components. ‘The connections that exist between the m elements of the network will impose constraints on components of the vectors a and b. Namely, where a connection exists, there will be an equality established between incident and output waves. This is demonstrated by the waves at a single connection shown in Fig. 3.1(b). These waves must satisfy ag=bj ay = dy. (3.3) Similar relations are apparent at every connection. This allows construction of a connection matriz I’ satisfying b=Ta (3.4)(b) Figure 3.1 An aggregate network, Snee, consisting of many interconnected components, each characterized by scattering parameters, Schematic diagram of the network (a), and detail showing an incident and output wave for components S; and Sj at a connection (b).35 The components of I’ are all zeros, except those entries that correspond to a connection, and have a value of unity. Application of (3.2) and (3.4) gives the solution for the output wave vector [1] b=T(P-s)c. (3.5) The composite network is therefore solved solely in terms of the component S- matrices and knowledge of the interconnections. This is a brute-force method. The (I — S) matrix, albeit sparse [2], can be quite large, and its inversion leads to long computation times. The solution (3.5) comes in terms of the composite wave vector b, giving output waves for the aggregate network, and also those for each component within the network, although the latter are generally of no concern. In addition, (3.5) cannot be applied directly to the source free (¢ = 0) case. This type of method may be used to perform noise analysis, as recently proposed by Dobrowolski [3]. These problems, however, inspire the need for an alternative. Improvement is possible through application of connection methods. The idea is to divide the network to be solved into subcircuits to simplify the overall calculation. Such an approach was first described by Murray-Lasso [4] using indefinite admittance matrices, and has also been described using wave variables in the absence of sources [5]. The first step is to separate internal and external variables of the network. In Fig. 3.1(a), connections internal to the network Snet are shown with dotted lines. Those connections external to the network, and accessible by terminals, are shown in solid lines. Simplification is obtained by avoiding outright calculation of internal variables. The process begins, as before, with tableau equation (3.1). Next, those waves which are external to the network are separated from those internal. This results in the partitioned form of (3.2): ()=(8 &)(5) + (8) e36 where the external waves are denoted by subscript ¢, and those internal by sub- script i. The topology of the network is based on connections between internal elements, so the connection matrix I may be redefined using bj =Tay. (3.7) As before, I’ designates equality between internal incident and outward waves, so its elements are either 1 or 0 for components with consistent definitions of nor- malizing impedance. Desired is the scattering matrix for the aggregate network Snet defined as the ratio of external incident and outward deterministic waves, and the equivalent source wave for the network Cnet that satisfy: = Snet Ae + Cnet (3.8) This is sufficient information to calculate the source free S-matrix for the net- work, Setting the source terms to zero and solving equations (3.6)-(3.8) yields the connection formula: Snet = See + Sei(T — Sii)“*Sie- (3.9) Solving the problem with sources included is accomplished by defining cnet as the waves that exit the network in the absence of input: (3.10) leading to cnet = [1 | Sai(P — 8u)"] () (3.11) where I is an identity matrix with order equal to the number of external ports, T is the same connection matrix defined by (3.7), and the bar (|) denotes matrix augmentation. If the source terms ce and cj are due to noise, then the solution must be put in terms of a noise wave correlation matrix. Using the partitioning37 method of (3.6), a noise wave correlation matrix may be defined in terms of cock Cael _ __}. (3.12) aes Ge This matrix is filled with values from the known noise wave correlation matrices of the components. The connection formula for the new correlation matrix is internal and external noise waves found by applying Cnet = CnetChey and (3.11) Caet = [1 | Sei(T -Su)“*] ©. [1 | Sei(T -Sy)?]!. (3.13) ‘This result, combined with (3.9), gives a complete noise and signal character- ization of the aggregate network. Several improvements have been made over the approach leading to (3.5). A direct calculation of the new scattering matrix is now possible using (3.9). In addition, the order of the matrix to be inverted has been reduced by an amount equal to the number of external ports. A very dramatic advantage also comes from the similarities in the expressions for Snet and Cyet- Both share the (T — $j)~! matrix in their calculations. It is this matrix inversion that requires most of the computational effort. Surprisingly, in approximately the same time required for performing a lone signal analysis, both signal and noise analyses may be performed.33 3.2 Network Reduction by Subnetwork Growth Unfortunately, for a circuit comprised of many components, the computer time and memory required to perform even the simpler (I— $;;) matrix inversion can be prohibitive. It does tend to be a sparse matrix, however, so appropriate methods are in order. General methods have been described elsewhere [2,6], yet an alternative sparse matrix approach is available for this specific problem. Known as subnetwork growth [1], it involves a repeated application of connec- tion formulas. Computation time is reduced by building up large networks by connecting smaller subnetworks together, two at a time, reducing the size of the matrix to be inverted. In the simple subnetwork growth procedure described here, only one connection is made at a time, reducing the size of the (I — Si) matrix to a 2-by-2. This process is illustrated in Fig. 3.2 for a branch-line coupler. ‘The coupler is a network comprised of eight parts: four quarter-wave transmis- sion lines, and four tees. These are labeled A through H in the figure, and the external ports are numbered from 1 to 4. In the first step, the parts are joined in pairs forming four three-port subnetworks. Next, these three-ports are com- bined in pairs forming the two four-ports, ABCD and EFGH. Finally, the two four-ports are joined to make the complete branch-line coupler. Subnetwork growth for deterministic analysis has been studied by Monaco and Tiberio [7] and Filipsson [8]. Since no intermediate results are stored, the method has a fundamental difficulty in that the entire process must be repeated for multiple analyses. However, in a comparison to a more standard sparse ma- trix method that included pre-processing for fast re-analysis, subnetwork growth was shown to be superior in speed for analyses numbering less than 120 [5]. Ex- ecution time is sensitive to the order in which the joints are made, so a search for the connection that results in a new subnetwork with the smallest number of parameters is needed.