Technical Report #USC 02-0511 Electronic Noise Characterization − Part I: System Concepts and Theory Dr. John Choma Professor of Electrical Engineering Scholar in Residence, Raytheon Space and Airborne Systems Electronics Center University of Southern California Ming Hsieh Department of Electrical Engineering University Park: Mail Code: 0271 Los Angeles, California 90089–0271 213–740–4692 [USC Office] 626–915–0944 [Home Fax] 818–384–1552 [Cell] johnc@usc.edu ABSTRACT: This report provides detailed discussions of the system concepts that mathematically underpin a characterization of electronic noise phenomena in electronic systems and circuits. Noise generated by thermal effects in two-terminal resistances, shot effects in active elements, and flicker noise phenomena exuding a generally troublesome inverse frequency dependence are studied at some depth. The analytical means for assessing the impact that these parasitic noise sources exert on the idealized (noiseless) properties of commonly encountered active networks receive considerable attention. This attention is synergistic with the fundamental goal of forging generalized, but mathematically tractable, circuit design strategies that mitigate, or at least minimize, noise-induced degradation of observable circuit performance. July 2011 Technical Report #02-0511_R1 1.0. University of Southern California Viterbi School of Engineering Choma INTRODUCTION The study of electronic noise in high performance analog integrated circuits requires a consideration of at least three issues. The first of these issues is the identification of the principle noise sources in an electronic circuit. To this end, we shall address the three most common sources of electrical noise. The first of these noise sources is thermal noise −often called Johnson noise or Nyquist noise−[1]-[2], which materializes in virtually all passive conductors. Unlike most other noise sources, thermal noise does not require that the afflicted device conduct current conduction. A second noise source is shot noise, which is evidenced whenever a device conducts a nonzero static current and presents a potential barrier that transported charges must overcome to sustain the directed current. Finally, flicker noise is principally observed at relatively low frequencies. Like shot noise, flicker noise requires a directed current flow and in general, the amount of flicker noise increases with increases in current level. An especially troublesome aspect of flicker phenomena is that the “low” frequency passband over which it can be a significant, if not dominant, source of electrical noise in an electronic system can extend to distressingly higher frequencies as device unity gain frequency capabilities expand. Popcorn noise (also termed burst noise), which seemingly derives from metal ions that contaminate a semiconductor body, is not addressed in this report owing to the unavailability of satisfying analytical formulae that relate burst phenomena to relevant physical constraints or measureable electrical parameters. Fortunately, popcorn noise, which has been observed in gold-doped bipolar junction transistors, is generally manifested only sporadically and only at frequencies that are far below the signal frequencies associated with broadband and radio frequency (RF) bandpass amplifiers. Finally, a consideration of excess noise, which is a kind of flicker noise effect exhibited by carbon composition resistors, is similarly spurned in subsequent discussions[3]. The second noteworthy issue implicit to this study of electrical noise is the designoriented analysis promoting the phenomenological understanding of noise properties that underpins prudent low noise design. At possible risk of oversimplification at this early juncture of our work, the dominantly undesirable effect of electrical noise phenomena is that noise voltages and currents coalesce with applied signals to the point that signal amplitudes can become comparable to, or even dwarfed by, noise. The analyses undertaken herewith therefore focus on theoretic methods and practical design strategies that preclude signal obscurity in the face of electrical noise, along with ensuring the reliable detection and processing of signals immersed in inherently noisy environments. We shall demonstrate that this signal detection and processing requires judicious biasing and other optimal design strategies that minimize the effective input noise voltages and currents applied to an electronic system. To this end, we shall introduce generalized design strategies that promote lowering the noise floor of a network to a level that ensures reliable observability and efficient processing of small input signals. The final issue is addressed by the second part of this report. This vital engineering issue is the documentation and application of the circuit level noise models that underpin our analytical endeavors. To this end, we shall concentrate on noise models for the passive resistor, the PN junction diode, and the MOSFET. We shall then exploit these models to assess the noise characteristics and properties of a few commonly encountered amplifiers. Electronic Circuit Noise, Part I_R1 -2- July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma TIME DOMAIN REPRESENTATION[4] 2.0. Electrical noise is a random voltage or current waveform. The amplitudes of these noise voltages and currents typically subscribe to a Gaussian distribution and therefore, they cannot be quantified at specific times. Because a clarion analytical representation of noise in the time domain is not practicable, we have few engineering options but to resort to such statistical characterization metrics as the mean square value of noise and its companion, the root mean square (RMS) value. We shall also deem it profitable to exploit such additional engineering performance metrics as the signal -to- noise ratio (SNR), the noise figure or noise factor, and the correlation coefficient associated with the algebraic combination of noise waveforms. In the subsections that follow, we examine these metrics and their implications for the special, but commonly encountered, case of a noise voltage or current having zero average value. We proffer that a zero average noise value is common because if a waveform is inherently random in nature, the subject waveform has equal probabilities of being positive or negative over time. 2.1. ROOT MEAN SQUARE VALUE OF NOISE The root mean square (RMS) value, Erms, of a noise voltage waveform, symbolized abstractly in the time domain as en(t), is T E rms 1 e 2 (t)dt T n 0 e 2 (t) , (1) n where e 2 (t) , which denotes the mean-squared value of en(t), is nonzero even if en(t) projects zero n average value. We point out in the interest of clarity that throughout this report, we shall invoke the symbol, “E,” to designate the RMS value of a random voltage waveform. Moreover, “v” and “i” designate time deterministic voltage and current, respectively, while “e” and “j” are their respective random, or noise, counterparts. We note from (1) that the square of an RMS voltage returns the mean square value of the subject voltage. In (1), T is a sufficiently long averaging period. If the random process underlying en(t) is ergodic1, the value, T, selected for the integration process is unimportant. In actual laboratory practice, however, large T generally produces more meaningful mathematical results than does a small averaging interval. In the interest of completeness, a noise current waveform, in(t), has an RMS current value, Jrms, given by J rms 1 T T 2 jn (t)dt 0 j 2 (t) . (2) n Once again in the interest of analytical clarity, symbol “J” designates the RMS value of noise current, whence the square of the RMS current value is the mean square noise current. For all practical purposes, the RMS voltage or current stipulates the noise power associated with a random voltage or current. Indeed, the square of the RMS voltage developed across a 1-ohm resistance and the square of the RMS current conducted by the same 1-ohm resis- 1 Ergodicity refers to a random process in which every sequence or sample of the time domain variable is equally representative of the entire time domain waveform. Electronic Circuit Noise, Part I_R1 -3- July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma tance can be thought of as the normalized noise power, Pn, dissipated in the resistance. In particular, we note that P E n 2 rms 1 or very simply, P E n 2 1 , J rms (3) 2 2 J , rms rms (4) where it is understood that Pn is in units of watts despite being numerically equal to the square of either RMS noise voltage or RMS noise current. When expressed in terms of the square of the RMS voltage developed across a resistive branch, the normalized noise power scales in proportion to branch conductance. On the other hand, Pn scales with branch resistance when power is expressed in terms of the squared RMS current flowing through the resistive branch. 2.2. SIGNAL TO NOISE RATIO AND NOISE FIGURE The signal to noise ratio, SNR, at a stipulated circuit port is commonly (but not universally) expressed in units of decibels (dB) and is given by 2 V 2 V v (t) rms 20 log rms , (5) SNR 10 log 10 log E 2 E e 2 (t) rms rms n where the port of interest is understood to support a time-deterministic signal, v(t), which is immersed in a noise voltage. The normalized power of the time deterministic signal is V 2 , while the normalized noise power implicit to the signal prevailing at the port of interest is rms E2 . rms It is intuitively clear that we wish the SNR to be as large as possible. For example if the RMS signal voltage established at a given network port is 5 mV and the corresponding RMS noise voltage is 1 μV, the SNR is a laudable 74 dB. But if Vrms = Erms, SNR = 0 dB, which suggests the improbability of discriminating between signal and noise voltages because neither signal nor noise is dominant at the subject port. Of course, if the SNR is not expressed in units of decibels, it assumes a simple numerical value, and (5) becomes 2 V V2 rms rms , SNR 2 E E rms rms which is effectively a ratio of signal power -to- noise power evidenced at the input port. (6) The SNR is a first cousin to the noise factor, F of a network. A demonstration of the relationship between signal -to- noise ratio and noise figure best derives from a consideration of the linear network abstraction in Figure (1). In this diagram, a signal source applied to the linear network manifests an RMS signal voltage at the input port of Virms and a resultant signal response at the output port of Vorms. If the linear network is an active circuit, input/output (I/O) gain is plausible so that Vorms > Virms. But all networks are unavoidably contaminated by noise, which means, among other things, that the reliable detection and processing of the applied input signal is rendered potentially problematic. If, in the present case, the RMS noise voltage measured at the input port and due exclusively to noise associated with the applied input signal is Eirms, the resultant signal -to- noise ratio evidenced at this input port is Electronic Circuit Noise, Part I_R1 -4- July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma Input Signal Source Virms Eirms * Output Port Linear Network * Eorms Vorms Load Input Port Figure (1). Symbolic depiction of a linear network to which an input signal of RMS value Virms is applied to the input port and a corresponding signal of RMS value Vorms results at the output port. Both input and output ports are contaminated by noise voltages (abstracted here as “*” within the voltage generator symbol) whose respective RMS values are Eirms and Eorms. 2 V irms SNR , i 2 E irms (7) where we hope that SNRi is substantively larger than unity; that is, we wish the signal to dwarf the noise at the input port, as opposed to noise blurring the signal identified for processing. In response to the source-derived input noise measured by voltage Eirms and noise generated within the network undergoing investigation, an RMS output port noise voltage of Eorms is observed, thereby presenting an output port signal -to- noise ratio of 2 orms SNR . o 2 E orms V (8) Then, the noise factor of the linear network undergoing investigation is defined numerically as SNR i , F (9) SNR o or in decibel units (whereupon it becomes known as the noise figure) as SNR i . (10) F 10 log SNR o Thus, the noise figure is little more than the decibel value of the noise factor. Note, that because the RMS signal and noise voltages appear as squared quantities in the signal -to- noise ratio expressions of (7) and (8), these signal -to- noise ratios are given as ratios of mean squared (or square of RMS quantities) signal -to- noise quantities. Under ideal circumstances, which are observed only within the confines of hallowed academic halls of ivy, the linear network under present consideration generates no noise in and of itself. Under such a metaphysical circumstance, the only noise observed at the output port is the noise incident with the input port, multiplied, or amplified, by the I/O transfer function, or gain in the case of a linear active configuration. But since the transfer function multiplying the input noise is the same as that which multiplies the input port signal, the signal -to- noise ratio, SNRo, observed at the output port is necessarily equal to SNRi, whence F = 1 (or 0 dB). Electronic Circuit Noise, Part I_R1 -5- July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma In contrast with the foregoing ideal consideration, if noise is generated within the linear network, the net output noise is larger than the amplified input noise, which renders SNRo (which is inversely proportional to noise level) smaller than SNRi. Accordingly, F exceeds 0 dB for practical networks while for noiseless circuits and systems, noise figure F assumes its idealized value of 0 dB. It should be clearly understood that the ideal situation for which F = 0 dB does not imply zero noise level at the output port. Instead, it implies only that the noise contributed by the linear network to the net noise level manifested at the network output port is negligible in comparison to the output noise level incurred exclusively because of the amplified input noise. 2.3. CORRELATION COEFFICIENT A commonly encountered situation in the noise analysis of active or passive structures is the case when two noise sources algebraically superimpose with one another. Figure (2a) illustrates this superposition issue for noise voltages en1(t) and en2(t), whose respective RMS values are E1rms and E2rms. Figure (2b) is the counterpart depiction of two superimposed noise currents, jn1(t) and jn2(t) possessed of respective RMS values of J1rms and J2rms. If the two noise voltages sum to produce the resultant noise response, en(t), as suggested in Figure (2a), its RMS value, Erms, derives from (11), where we have used (1) as the pertinent mathematical basis: jn(t) * en1(t) en(t) jn1(t) jn2(t) * en2(t) jn(t) (a). (b). Figure (2). (a). Superposition of two noise voltages. No polarity is assigned to the resultant terminal noise voltage, en(t) because of the random nature of the two noise voltages, en1(t) and en2(t). (b). Superposition of two noise currents. No direction is ascribed to the flow of the resultant net current, jn(t) because of the random nature of the two noise currents, jn1(t) and jn2(t). T T 2 1 1 2 2 2 2 E e (t) e (t) e (t) dt e (t) e (t) 2e (t)e (t) dt . (11) rms n n1 n2 n1 n2 n1 n2 T T 0 0 The integral averaging of the first term on the far right hand side of this expression produces the mean square value of noise voltage en1(t); that is, 1 2 2 e (t) E n1 1rms T T 2 en1(t)dt . (12) 0 Similarly, the mean square value of noise voltage en2(t) is 1 2 2 e (t) E n2 2rms T T 2 en2 (t)dt . (13) 0 In (11), we introduce the correlation coefficient, κ[5], such that Electronic Circuit Noise, Part I_R1 -6- July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma T 1 e (t)e (t)dt T n1 n2 0 E . E (14) 1rms 2rms It can be shown that −1 ≤ κ ≤ 1, where κ = 0 implies that the two noise signals are uncorrelated, which means that the subject noise voltages derive from mutually independent physical processes. On the other hand, κ = ±1 asserts maximal, or full, correlation between en1(t) and en2(t). Equation (11) now becomes E 2 2 2 E E 2 E E . rms 1rms 2rms 1rms 2rms (15) We see that for two uncorrelated noise voltages, for which κ = 0, the mean square value of the noise voltage sum is simply the sum of the individual mean square noise voltages. On the other hand, the mean square value of the sum of two fully correlated noise voltages, which engender κ = ±1, is E2 rms E2 1rms E2 2rms 2E E 1rms 2rms E1rms E2rms 2 , (16) where the plus sign applies to the case in which en1(t) and en2(t) correlate in phase with one another, while the minus sign pertains to two noise signals that correlate anti-phase with one another by. Of course, for the case of two noise currents, as depicted in Figure (2b), the mean square value of the noise current sum is, by analogy to (15), J 2 2 2 J J 2 J J , rms 1rms 2rms 1rms 2rms (17) where it is understood that the correlation coefficient, κ, is now 3.0. 1 T T jn1(t) jn2 (t)dt 0 J J . (18) 1rms 2rms FREQUENCY DOMAIN ANALYSIS[4] Unlike periodic signals that deliver power at only their fundamental frequency and possibly discrete other frequencies, noise voltages and currents have their power distributed continuously over a frequency spectrum. This frequency domain distribution of noise power encourages a characterization of a noise voltage, say en(t), by its noise voltage spectral density, SV(f), which projects, as a function of frequency, the measured or theoretically discerned mean square voltage (or normalized power) delivered to an arbitrary load -per- unit of frequency. Thus, for a noise voltage, the dimensional unit of its noise spectral density is volts2/Hz (in reality, mean square volts/Hz). For a noise current, say jn(t), its noise spectral density, say SI(f), has units of amps2/Hz ((in reality, mean square amps/Hz). The immediate sidebar to the noise spectral density is that it offers an ostensibly convenient frequency domain alternative to the time domain calculation of mean square value projected by (1) -through- (4). In particular, the mean square value (or square of the RMS value) of noise voltage en(t), which postures a noise spectral density of SV(f), is given by Electronic Circuit Noise, Part I_R1 -7- July 2011 Technical Report #02-0511_R1 1 2 E rms T T 0 University of Southern California Viterbi School of Engineering Choma 2 2 e (t)dt e (t) n n SV (f)df (19) . 0 Of course, the current counterpart to the voltage expression in (19) is 1 2 rms T J T 2 n 2 n j (t)dt j (t) 0 S I (f)df (20) , 0 where SI(f) is the spectral density of noise current abstracted in the time domain as jn(t). Because of the integrations implicit to the preceding two relationships, we note that assuming finite noise spectral densities, normalized power, for which mean square voltage or mean square current is a measure, is not delivered by any noise voltage or noise current at a single, specific frequency. Rather, such noise power is delivered only over a band, or range, of frequencies. A simple demonstration of the last contention proves instructive. Assume that the noise spectral density of a certain noise voltage is S (f) K e V f f o o , (21) where Ko is a constant whose dimension is volts2/Hz, and fo is a constant frequency parameter. Using (19), we see that the mean square value of this noise voltage is 2 E rms K oe f f o df K f e f f o o o 0 K f o o 0 . (22) Alternatively, we may inquire as to the normalized power delivered by the noise voltage over a frequency passband extending from f1 -to- f2, where f2 > f1. Such an inquiry is meaningful, for example, if the noise interfaces with a sharply tuned bandpass filter. Then the normalized noise power, say Pn, delivered in the frequency interval from f1 -to- f2 is f P n 2 f K e f f o o df K f e o o f f o f f 1 2 1 f f f f K f e 1 o e 2 o o o (23) f f f1 f o 2 o , e e rms where we have applied (22) to our band limited noise power calculation. Observe in this example that Pn = 0 if f2 = f1; that is the noise power delivered at a single frequency is zero. E2 3.1. INPUT/OUTPUT NOISE PROCESSING Consider the case in Figure (3a) of a noise voltage, eni(t), applied to the input port of a linear network whose (I/O) transfer function is H(s). We assume that the linear network itself is noiseless or, more pragmatically, that we wish to examine only the filtering effect that the considered linear network has on the applied input noise. If the input noise voltage has a noise spectral density of SVi(f) and the resultantly generated output noise, eno(t), displays a noise spectral density of SVo(f), we proffer S Vo (f) H j2πf Electronic Circuit Noise, Part I_R1 2 (24) S (f) , Vi -8- July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma which exploits several basic principles. The first of these nuances is that since the spectral density of a noise voltage projects mean square volts/Hz, the multiplier linking the input noise spectral density to its resultant output noise spectral density is the squared magnitude of the transfer function. This requirement follows from the fact that a mean square voltage response can never be negative, nor can it ever be a complex number. Using the magnitude of a transfer function, as opposed to a mere transfer function in the complex frequency plane, certainly obviates negative and complex numerical offerings in (24). Squaring the transfer function magnitude follows immediately from the fact that the noise spectral density is focused on mean square voltage -perhertz. Embodied within this first principle is the second principle, which is that noise spectral density observations apply only to steady state operation. Assuming that the considered linear network is stable, this means that the complex frequency argument, s, in the transfer function representation of Figure (3a) must be replaced by its steady state counterpart, s = jω = j2πf. Finally, we note that (24) is divorced of any I/O phase information. This circumstance reflects our focus on purely random phenomena that renders a practical time domain representation intractable, thereby obfuscating time domain shifts between applied noise input and resultant noise output response. The immediate ramification of (24) is a mean square output noise given by2 eni(t) H(s) eno(t) en1(t) H1(s) en2(t) H2(s) en3(t) H3(s) (a). + eno(t) (b). Figure (3).(a). An input noise voltage, eni(t), applied to a linear network (possibly a filter) whose I/O transfer function is H(s). The input noise voltage has a noise spectral density of SVi(f). (b). The application of three uncorrelated noise voltages to three individual subcircuits whose noise output responses are algebraically summed to produce the indicated output noise voltage, eno(t). T 1 2 2 2 2 E e (t)dt e (t) S (f)df H j2πf S (f) df . orms no Vo Vi T no 0 0 0 (25) In simple street talk, the output noise spectral density of the system illustrated in Figure (3a) is the input noise spectral density scaled by the squared magnitude of the frequency domain I/O transfer function. And since the mean square noise voltage is the noise spectral density inte2 We just as easily could have formulated the following conclusion in terms of input noise current and indeed, any combination of input and output noise voltages and currents. Electronic Circuit Noise, Part I_R1 -9- July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma grated over all positive real frequencies, the output mean square noise voltage is the frequency integral over all non-negative frequencies of the product of the input noise spectral density and the squared magnitude of the system transfer relationship. Consider now the more complicated system shown in Figure (3b) wherein three uncorrelated noise voltages, en1(t), en2(t), and en3(t), are applied independently to three linear subcircuits whose respective transfer functions are H1(s), H2(s), and H3(s). The output responses of these three subcircuits are algebraically summed to forge the output noise voltage, eno(t). Because the considered noise sources are presumed uncorrelated, the mean square value of the superposition of their noise output responses is simply the sum of the individual mean square output responses. Thus, the noise spectral density, SVo(f), of the output noise voltage response in Figure (3b) derives straightforwardly as S Vo (f) H i j2πf i 2 S (f) , (26) Vi where we have borrowed from (25) and (15) and have identified Svi(f) as the noise spectral density of the ith input noise voltage, eni(t), in the diagram of Figure (3b). We have also allowed an open summation over running index i in (26) since it is clear that any number of input noise voltages can interact with the corresponding number of subcircuits, which effectively act as noise filters. It follows that the net mean square output noise voltage (or square of the output RMS noise voltage) is, by way of (25), 2 E orms 3.2. SVo (f)df 0 i 0 H j2πf i 2 S (f) df . Vi (27) NOISE EQUIVALENT BANDWIDTH As we have demonstrated in the preceding section of material, the frequency domain computation of the mean square noise voltage developed across the output port of a linear network derives from integrating over all non-negative frequencies the output noise spectral density, SVo(f). Since this output spectral density embodies the product of the input noise spectral density, SVi(f), and the squared magnitude, |H(j2πf)|2, of the applicable network transfer function, the requisite integration over frequency is likely to be challenging, particularly if the transfer function embodies a number of significant poles and/or zeros. This cumbersome computation motivates a search for simpler computational methods that apply to most, if indeed not all, noise analysis undertakings. To the foregoing end, consider the diagram in Figure (4a), which depicts a linear network to which a noise voltage, eni(t), is applied to the network input port. If the noise spectral density of eni(t) is SVi(f), as is suggested in the diagram, its exclusive contribution to the resultant noise spectral density, SVo(f), manifested at the network output port is S Vo (f) H j2πf 2 (28) S (f) , Vi where, of course, H(j2πf) represents the I/O network transfer function in the steady state. By concentrating exclusively on the input noise, as opposed to including the effects of noise sources generated within the network, SVo(f) in (28) effectively invokes the simplifying presumption of a noiseless network. Equivalently, we can state that the foregoing simplification corresponds to a network boasting 0 dB noise figure (or unity noise factor). It therefore follows that the mean Electronic Circuit Noise, Part I_R1 - 10 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma square value of the generated output port noise in the system of Figure (4a) remains pinned by (25). Input Port SVi (f) eni (t) Linear Network, H(j2f) Output Port SVo (f) Input Port SVi (f) eno (t) eni (t) |H(j2f)| Brick Wall Network, HB(j2f) Output Port SVoeff (f) enoeff (t) |HB(j2f)| Hm Hm 0.707Hm f B f (a). f (b). Figure (4). (a). A linear network that generates an output noise voltage response, eno(t), which has a noise spectral density of SVo(f), to an applied input noise voltage, eni(t), whose noise spectral density is SVi(f). The subject linear network, whose noise is tacitly ignored, is presumed to be a lowpass structure having a 3-dB bandwidth of B (in Hz). (b). The input noise voltage of (a) applied to an ideal, noiseless, “brick wall” filter that, over a frequency passband of Δf (in Hz), delivers constant gain magnitude Hm that is identical to that of the lowpass structure in (a). We now replace the actual network used in Figure (4a) by the mathematical −and notoriously idealistic and physically unrealizable− “brick wall” system in Figure (4b), whose transfer function, HB(j2πf), evokes a magnitude defined by H , for 0 f Δf m (29) H j2πf , B 0 , for f Δf where Hm is identically the zero, or low, frequency gain magnitude of the original network in Figure (4a). If the input noise applied to the brick wall system is the same as that applied to the original network, the noise spectral density, SVoeff(f), manifested at the brick wall output port follows as (f) H B j2πf 2 S (f) H 2 S (f), for 0 f Δf m Vi . 0 , for f Δf Thus, the mean square value of the brick wall output port noise voltage, vnoeff(t), is S Voeff e2 (t) E 2 noeff oeffrms Vi 0 S (f)df H 2 Voeff m (30) Δf 0 S (f) df , Vi (31) where we have made use of (30) and the fact that gain parameter Hm is a frequency invariant constant. Electronic Circuit Noise, Part I_R1 - 11 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma Now, let us choose the foregoing Δf metric so that the net mean square noise voltage at the output port of the brick wall filter is identical to the mean square output noise voltage generated in the original system; that is f 2 2 H S (f) df S (f)df H j2πf S (f)df , (32) m Vi 0 o Voeff o which produces 2 f H j2πf S (f) df S (f)df . Vi Vi H 0 Vi (33) m o When frequency metric Δf is chosen in accordance with (33), it is termed the noise equivalent bandwidth of the original system earmarked for an electrical noise characterization. The relationship in (33) postulates that frequency metric Δf can be selected so that the two filters in Figure (4) deliver the same mean square output noise voltage for the same amount of applied input noise. Specifically, output noise equality results if the area enclosed by the input noise spectral density, SVi(f), over a frequency range spanning from zero -to- Δf, is equal to the area enclosed over all frequencies (0 -to- ∞) by the product of SVi(f) and the squared magnitude of the original network transfer function normalized to its zero frequency gain value. We note with more than passing interest that the noise equivalent bandwidth, Δf, is the width of brick wall filter passband; that is Δf is the passband of the brick wall filter. Although (33) postures elegant mathematics, its pragmatic execution for a generalized transfer relationship and for any generalized input noise spectral density is potentially cumbersome. It is inevitably a daunting undertaking unless we elect to assume a constant noise spectral density. Constant noise spectral density, which suggests that the mean square noise -per- unit of frequency is independent of frequency, typifies a type of electrical noise that is called white noise. As it turns out, thermal noise in traditional integrated circuit resistances and shot noise in semiconductor devices very closely approximate sources of white noise. And since thermal and shot noise phenomena dominate the noise landscape of active circuits, especially at moderately high -to- high signal frequencies, constraining SVi(f) to a constant reflects sensible circuit design engineering. With SVi(f) constant, (33) collapses to the considerably simpler expression, f o H j2πf H 2 df . (34) m This relationship stipulates the constraint under which frequency metric Δf can be selected so that the two filters in Figure (4) deliver the same mean square output noise voltage for the same amount of applied input white noise. Numerous lowpass, high performance, electronic circuits and systems are designed to deliver maximal 3-dB bandwidth, subject to the constraint of no, or at least minimal, peaking of the frequency response over the network passband. When no response peaking is evident, such broadband systems are said to deliver a maximally flat magnitude (MFM) response. To this end, a common design target is the Butterworth lowpass system whose I/O transfer function subscribes to the relationship, Electronic Circuit Noise, Part I_R1 - 12 - July 2011 Technical Report #02-0511_R1 H j2πf 2 University of Southern California Viterbi School of Engineering Choma H2 m (35) , 2n f 1 B where we remind that Hm is the low frequency magnitude of the transfer relationship, while n, which is the order (number of poles) of the considered network, is an integer that is at least as large as one. For any value of n, B in (35) represents the applicable 3-db bandwidth. If we substitute (35) into (34) in an attempt to gain a basic perspective as to the relationship between the noise equivalent bandwidth and the actual 3-dB bandwidth for signals, we find that for the Butterworth MFM filter Δf 2n (36) . B sin 2n For n = 1, which corresponds to a dominant pole, or ideally single pole, circuit, Δf/B = π/2, which indicates that the noise bandwidth is more than 50% larger than the signal 3-dB bandwidth. The fact that Δf exceeds B substantially is not surprising in that a single pole network has a relaxed, 20 dB/decade magnitude response roll off at frequencies in the neighborhood, and in excess, of the network 3-dB bandwidth, B. Consequently, noise at frequencies larger than the network 3-dB bandwidth can be amplified to some degree to deliver a potentially large mean square output noise over wide high frequency passbands. In contrast, no noise is delivered by the brick wall network for frequencies larger than Δf. Accordingly, Δf must be suitably larger than B to account for noise delivered at high signal frequencies to the output port of the practical single pole system. The fact that Δf always exceeds B, supplemented by the fact that since (31) confirms a mean square output noise that is directly proportional to Δf when the applied noise is white, discourages us from exercising the design heroics to achieve circuit bandwidths that dramatically exceed the required system bandwidth. In particular, since Δf is proportional to B, larger bandwidths produce progressively larger normalized output noise power or mean square output noise voltage or current. While (36) is strictly applicable to only a Butterworth MFM circuit or system, it generally serves as a reasonable design guideline for choosing a circuit or system bandwidth, B, which suitably restricts the total mean square noise (voltage or current) developed at the network output port. Indeed, the case of N = 1, which we have just considered, applies to any dominant pole circuit or system that exudes no finite frequency zeros. In the interest of completeness, N = 2 in (36) establishes Δf/B ≈ 1.11, which suggests a noise equivalent bandwidth that is 11% larger than the signal bandwidth. For 3 ≤ N ≤ 9, less than about 4% error accrues by approximating (36) with Δf/B ≈ π/3. In a word, the noise equivalent bandwidth of most practical circuits and systems abide by the design-oriented constraint, Δf , for 2 N 9 . (37) 2.83 B 3 We note that Δf converges toward bandwidth B for progressively higher order networks. In light of our previous frequency response observations, this situation is to be expected since high order networks exude robust frequency rates of magnitude attenuation at high frequencies, thereby leaving only an anemic gain magnitude with which to process input noise at high frequencies. 3.3. NOISE SOURCES IN THE FREQUENCY DOMAIN Thermal noise generated in resistors, or conductors in general, as well as shot noise in electronic devices are very good approximations of commonly encountered white noise. On the Electronic Circuit Noise, Part I_R1 - 13 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma other hand, flicker noise, which is often called pink noise, is pervasive of many active semiconductor devices. It is an excellent example of a commonly encountered noise source whose noise spectral density is not frequency invariant. Noise In Resistors 3.3.1. Noisy R R * enR(t) [SVR(f)] (a). R (b). Noiseless Noiseless Practical resistances, such as the resistance of value R (in ohms) given in Figure (5a), produce an open circuit noise voltage, enR(t), regardless of whether or not the resistors are connected in a circuit branch that conducts current. We depict this open circuit, or effective Thévenin, noise voltage in the noise equivalent model of Figure (5b), wherein enR(t) appears in series with the noiseless equivalent of resistance R. Of particular significance is the fact that the noise voltage spectral density, SVR(f), corresponding to enR(t) is given by jnR(t) [SIR(f)] (c). Figure (5). (a). A practical two-terminal resistor generating white thermal noise. (b). Thévenin type noise equivalent circuit of the noisy resistor in (a). (c). Norton type noise equivalent circuit of the noisy resistor in (a). S VR (f) 4 kTR , (38) where k is Boltzmann’s constant [1.38(10−23) joules/°K] and T is the temperature (in Kelvin degrees) of the subject resistor. To the extent that (38) correctly defines the noise voltage spectral density of resistor noise, which is often termed thermal noise because of the direct dependence of its noise spectral density on temperature T and the invariance of this spectral density with current level, resistor noise is seen as an example of white noise. This means that if resistance R is embedded in a circuit having a noise equivalent bandwidth of Δf, (19) and (31) suggest that the net mean square voltage that resistance R delivers to the circuit branch in which it is incident is e 2 (t) nR Δf SVR (f)df 4kTR 0 df 4kTRΔf . (39) 0 Thus, for example, a 1 KΩ resistance functioning at 27 °C (300.16 °K), generates noise voltage in the RMS amount of 2 f nR e 4kTR = 4.07 nV/ Hz (read as “4.07 nanovolts -per- root hertz”). Over a 1 GHz noise bandwidth, this computation implies a net RMS noise level of almost 130 μV! Instead of representing a practical resistor as a series interconnection of its noiseless counterpart and a noise voltage whose mean square value3 is given by (39), we can replace the 3 Remember that a mean square voltage is the square of its corresponding root mean square (RMS) voltage. Analogously, a mean square current is the square of its corresponding RMS current value. Electronic Circuit Noise, Part I_R1 - 14 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma Thévenin topology in Figure (5b) by the Norton noise equivalent circuit shown in Figure (5c). In the latter representation, the noise current, inR(t), has a mean square value of 2 e (t) 4kTRΔf 2 nR j (t) 4kTGΔf , nR 2 2 (40) R R where G is obviously the conductance equivalent of resistance R. Equation (40) also implies a noise current spectral density, SIR(f), given by S (f) 4 kTG , (41) IR whose frequency domain nature remains white. Noise In Junction Diodes 3.3.2. Unlike the passive resistor, the PN junction diode depicted in Figure (6a) generates an open circuit noise voltage, enD(t), only when it conducts a static, or quiescent, current, say IDQ, which naturally implies forward biasing of the device. When noise is functionally dependent on static current, it is termed shot noise, as opposed to resistive thermal noise, which is devoid of any current dependence. The Thévenin noise equivalent circuit of the diode is offered in Figure (6b), while Figure (6c) gives the Norton alternative to the Thévenin noise model of the diode. In Figures (6b) and (6c), rD, the small signal terminal resistance of the diode is given by IDQ Noisy rD (a). jnD(t) [SID(f)] rD enD(t) [SVD(f)] * (b). (c). Figure (6). (a). A forward biased diode generating noise that is functionally dependent on the quiescent current, IDQ, which it conducts. (b). Thévenin type noise equivalent circuit of the noisy diode in (a). (c). Norton type noise equivalent circuit of the noisy diode in (a). In both (b) and (c), rD represents the small signal resistance of the diode. r D n V j T I , (42) DQ where kT (43) T q is Boltzmann’s voltage, k is the previously introduced Boltzmann’s constant, T is the absolute temperature of the diode junction, and q is the magnitude of electron charge, which is 1.6(10−19) coulombs. Finally, parameter nj is the junction injection coefficient; it is typically very slightly greater than one but unless explicitly specified, its default value is routinely taken to be one. The diode noise voltage has a noise spectral density, SVD(f), of S (f) 2n kT r . (44) V VD j Electronic Circuit Noise, Part I_R1 D - 15 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma Since this spectral density is independent of signal frequency, diode shot noise comprises a second example of white noise that is commonly encountered in electronic circuits. It follows from (44) that the total mean square shot noise developed by a diode embedded in a network having a noise equivalent bandwidth of Δf is e 2 (t) 2n kT r Δf , nD j (45) D which is roughly one-half of the mean square voltage associated with a traditional passive resistance. Assuming room temperature operation of a semiconductor junction diode that is conducting 1 mA of quiescent current, rD in (42) is, assuming nj = 1, rD = 25.89 Ω. If the applicable noise equivalent bandwidth is 1 GHz, (45) gives a net RMS diode noise voltage of EDrms = 14.64 μV. Although this noise voltage is small, it is directly proportional to absolute diode junction temperature, which rises with increasing current densities at the diode junction. As in the case of the previously considered noisy resistor, the Thévenin noise model of the PN junction diode can be transformed into the Norton alternative displayed in Figure (6c). The Norton noise current, jnD(t), establishes a mean square value of 2 2n kT r Δf e (t) j D 2 nD j (t) 2qI Δf , nD DQ 2 2 r r D D (46) where we have adopted (45) and (42). For the numerical disclosures invoked in the diode noise voltage calculation, JDrms = 567.7 nA, which is of the same order of magnitude as the drain currents that typify a MOSFET biased in its subthreshold regime. However, resistance rD is typically of the order of only the low tens of ohms. Thus, the amount of noise current that a PN junction diode can supply to a load imposed across the terminals of the diode is small, assuming, of course, the likely circumstance of a load resistance that is significantly larger than rD. 3.3.3. Noise In MOSFETs In a MOSFET, such as the n-channel device appearing in Figure (7a) or its topologically identical p-channel counterpart in Figure (8a), three noise sources contribute to its overall noise response. These sources intertwine with the small signal model of the MOSFET to produce the noise and signal responses of the circuit into which MOS devices are embedded. The complete small signal model, with noise sources included, is not shown, but it will be addressed in a subsequent report when the noise characteristics of MOSFET technology amplifiers are studied. The first of the device noise sources is drain current noise, jnd(t)[6]. This noise is an amalgam of thermal noise manifested by that portion of the drain-source channel that is inverted and therefore acts as a tapered resistor and shot noise within the depleted region of the drainsource channel that is established when the MOSFET enters its saturation regime of operation. The second source of MOSFET noise is gate noise[7], which is caused by the thermal agitation of free carriers within the drain-source channel and, to a lesser extent, thermally noisy resistive gate material. Gate noise is most efficiently handled by incorporating into the gate lead of a MOSFET a series interconnection of an appropriate noise voltage eng(t), in series with a resistance, rg, While drain current and gate noises comprise reasonable approximations of white noise sources, the third noise component indigenous to a MOSFET is drain flicker noise[8]. This noise, which we model as a current, jnf(t), in shunt with the drain-source terminals of the transistor, is pink noise in that its mean square noise current varies inversely with frequency. Figure Electronic Circuit Noise, Part I_R1 - 16 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma (7b) offers the pertinent noise macromodel of the n-channel MOSFET. The p-channel MOSFET noise model is identical to that of the n-channel device and is offered in Figure (8b). In the subsections that follow, attention is focused almost exclusively on NMOS. However, the results disclosed apply equally well to PMOS with but minor and obvious notational changes. eng(t) Vds Vgs IdQ rg Noiseless * Noisy IdQ jnds (t) jnd (t) jnf (t) Vds Vgs (a). (b). eng(t) jnds (t) Noiseless * IdQ rg jnd s(t) Vds Vgs (c). Figure (7). (a). A conducting n-channel MOSFET that generates noise because of three distinct sources of internally generated noise. Biasing supportive of the quiescent drain current, IdQ, is not shown. Proper biasing requires a gate-source voltage, Vgs, larger than the threshold voltage, Vh and drain-source voltage Vds ≥ (Vgs − Vh). (b). Noise macromodel accounting for three internal noise sources for the device in (a). (c). The noise model of (b) with the two drain circuit noise components combined into a single drain noise current, jnds(t). Vsd Vsg eng(t) rg IdQ Noiseless * Noisy IdQ jnds (t) jnd (t) jnf (t) Vsd Vsg (a). (b). Figure (8). (a). A conducting p-channel MOSFET that generates noise because of sources of internally generated noise. Biasing supportive of the quiescent drain current, IdQ, is not shown. Proper biasing requires a source-gate voltage, Vsg, larger than the threshold voltage, Vh and source-drain voltage Vsd ≥ (Vsg − Vh). (b). Noise macromodel that accounts for three internal noise sources for the device in (a). 3.3.3.1. Drain Current Noise The drain current noise, jnd(t), is largely a thermal noise component when the MOSFET operates in its ohmic regime. But as operation draws toward pinch off and thence into saturation, Electronic Circuit Noise, Part I_R1 - 17 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma this current coalesces thermal noise phenomena with a modicum of shot noise. The mean square value of the drain current noise evidenced in a conducting MOSFET operated in its saturation regime is taken to be[6] j 2 (t) 4kT g Δf , nd m (47) where gm represents the forward transconductance of the transistor in saturation. It also happens to be the channel conductance, say gdo, in the ohmic regime of operation under the condition of a drain-source voltage, Vds, of zero. Specifically, W W (48) V V 2 C I , m c ox L gs h c ox L dQ where μc, in units of cm2/volt-sec, is the free carrier mobility in the channel (electron mobility, μn, for n-channel transistors and hole mobility, μp, for p-channel devices), while Cox is the density (in units of farads/cm2) of the gate oxide capacitance. Moreover, W/L is the gate width -tochannel length gate aspect ratio of the transistor, Vgs is the applied gate-source voltage, Vh denotes the threshold voltage, which is modulated somewhat by bulk-source voltage[9], and finally, IdQ represents the quiescent drain current flowing in the device In (47), 2 (49) 1, 3 with the understanding that the upper limit for γ applies to the ohmic domain of MOSFET operation, while the lower limit is suitable for the saturation regime. It should be noted that (47) effectively defines the mean square noise current of a simple passive resistance whose conductance value happens to be γgm. We should also point out that since hole mobility can be up to a factor of three or so smaller than electron mobility, conductance gdo, and hence the mean square value of the drain noise current, is smaller for PMOS than it is for a comparable NMOS transistor. g C 3.3.3.2. Gate Noise The mean square value of the gate noise voltage, eng(t), deployed in the gate circuits of the MOS transistors in Figures (7b) and (8b), is[7] e 2 (t) 4kT r Δf , ng (50) g where r 1 5g 1 W 5 C V V c ox L gs h 1 (51) , m W 50 ox I c T L dQ x εox is the dielectric constant of silicon dioxide [345 fF/cm], and Tox is the thickness of the gate oxide. Parameter δ, the gate noise coefficient, is typically assigned a value approaching δ = 4/3. In other words, δ ≈ 2γ when the considered MOSFET is biased in saturation. We observe that resistance rg, and hence the mean square gate voltage in (50), increases with longer channel lengths, narrower gate widths, and thicker gate oxides. It also increases with smaller drain currents. g 3.3.3.3. Drain Flicker Noise As we noted earlier, flicker noise, unlike white thermal or shot noises, produces a mean square voltage or current that varies as a nominal inverse function of frequency. Although flicker noise is known to occur in resistors, bipolar junction transistors, and even in vacuum Electronic Circuit Noise, Part I_R1 - 18 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma tubes, it is most prominent in electronic devices whose conduction mechanisms are principally determined by surface mechanisms and the properties of interfacial imperfections. Accordingly, MOSFETs, whose drain currents rely on the drift of free charge carriers in a thin, strongly inverted, drain -to- source channel that is induced at the surface interface between the gate oxide and semiconductor bulk, is vulnerable to flicker effects. This observation encourages the putative conclusion that flicker noise in MOSFETs is caused by imperfect semiconductor surfaces, charge entrapment at oxide-semiconductor interfaces, foreign impurities in oxide and bulk semiconductor volumes, and related other interfacial shortfalls that occur during routine device processing. The mean square value of the flicker noise current, jnf (t), in Figure (7b) is K f g 2 2 m Δf , (52) j (t) nf f WLC 2 ox where, as incorporated into (48), the saturation regime value of forward transconductance gm is[10] g 2 C W L I , (53) m c ox dQ with the understanding that carrier mobility μc assumes either its electron mobility value, μn, or its hole mobility value, μp, depending on whether the considered transistor is NMOS or PMOS, respectively. For NMOS, the flicker noise coefficient, Kf, is of the order of 5(10−27) coul/m2. In contrast, PMOS produces Kf in the neighborhood of (10−28) coul/m2; i.e. the flicker noise coefficient for PMOS is about 50-times smaller than that of NMOS. We recall that hole mobility is about three-times smaller than electron mobility. Thus, a PMOS transistor produces a flicker mean square noise that is as much as 150-times smaller than the flicker noise incurred by a comparably sized NMOS transistor conducting a drain bias current that is identical to that conducted by its PMOS counterpart. There are at least three other interesting sidebars to (52). The first of these luminescent observations derives from substituting the forward transconductance expression of (53) into (52) to arrive at the alternative mean square flicker current relationship of 2 c K f I dQ Δf . j 2 (t) (54) nf 2 L C f ox This result advances ominous undertones to channel length downsizing, as well as to large quiescent drain currents. In particular, we see that the mean square value of the drain flicker current is inversely proportional to the square of channel length and directly proportional to the static drain current. Thus, with all else being equal, a 45 nM MOSFET can exude as much as more than 8.3times the mean square value of drain flicker noise current than does a 130 nM transistor. A second interesting point is revealed by exploiting the facts that (1) the gate-source capacitance, Cgs, of a MOSFET operated in its saturated domain is 2 C WLC , (55) gs ox 3 and (2) the unity gain frequency, fT, of a MOSFET derives from Electronic Circuit Noise, Part I_R1 - 19 - July 2011 Technical Report #02-0511_R1 f T 2 C g University of Southern California Viterbi School of Engineering m gs C gd k g 3k g m m m m , 2 C 4 WLC gs ox Choma (56) where Cgd is the gate-drain capacitance of the transistor, we have adopted (55), and we have introduced parameter km to designate the capacitance divider, C gs (57) k . m C C gs gd Since the gate-drain capacitance in saturation collapses to only the capacitance associated with the gate oxide overlap with the drain implant, km is very nearly unity for a self-aligned gate MOSFET biased to operate in its saturation regime. If we now combine (56) and (52), we see that 2 4 f Δf 2 T j (t) WL K . nf f 3k f m (58) Thus, the mean square value of the drain flicker noise current is directly proportional to the gate area, which is WL. This behavior seems reasonable in view of the fact that larger gate areas enclose progressively larger volumes that can serve as sanctuaries for large amounts of trapped charges and greater volumes of ionic impurities. We also see that higher bandwidth devices, in the sense of enhanced fT, promote larger mean square flicker currents. Since increases in fT are generally realized through decreases in channel length L, this observation appears to resonate with the channel length conclusions we have drawn from (54). Moreover, since fT is nominally inversely dependent on gate-source capacitance, it can be reduced artificially by appending circuit capacitance across the gate-source terminals of the transistor. Obviously, this artificial reduction of the transistor unity gain frequency is sensible in only those applications that do not mandate large bandwidths. Clearly, the mean square flicker noise defined by (52), (54), or (58) is a monotone decreasing function of frequency, decreasing from very large values at low frequencies to ultimately inconsequentially small values at high frequencies. Because the flicker noise component, jnf(t), of drain current is shunted by, and therefore algebraically superimposes with, the thermal noise component, jnd(t), of the drain current noise, it is of interest to find the frequency at which the flicker noise component reduces to the thermal noise component. This frequency, say fc, is called the flicker noise corner frequency. It can be found simply by equating the mean square value of the flicker current in (58) to the mean square drain noise current in (47). The result is f 4WLK T f 3k m f c kT g 2 . (59) m Since the drain noise and drain flicker currents are uncorrelated, the mean square value of the net drain-source noise current, jnds(t), which we show in Figures (7c) and (8b), is the sum of the individual mean square values of the drain noise current and the flicker noise current. From (59), (58), and (47), this net mean square current is therefore expressible as f j 2 (t) 4kT g 1 c Δf . (60) nds m f Electronic Circuit Noise, Part I_R1 - 20 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma In the sense that potentially large flicker noise dominates the drain noise for frequencies f satisfying f < fc, (59) and its companion equation, (60), comprises arguably bad news for high performance device enthusiasts. In particular, wideband devices, which naturally boast large fT, have, by (59), large corner frequencies, fc. This state of affairs implies by virtue of (60), that signal frequencies f must be progressively larger to render f >> fc, and hence, a drain noise that is both reasonably small and nominally independent of frequency. The last point is highlighted graphically in Figure (9). In this figure, we plot the normalized mean square value of the drain current noise in (47), the normalized mean square value of the flicker noise component that derives from (58) and (59), and the normalized mean square value of the net drain-source current in (60). The drain current noise and flicker noise are assumed uncorrelated, and the normalization current factor for all three curves is (4kTγgmΔf). Note that in the figure before us, signal frequency f must be at least three-fold the corner frequency, fc, in order for flicker effects to be deemed inconsequential. Normalized Mean Square Noise Current 12 10 Net Drain-Source Noise 8 6 Drain Current Noise 4 Drain Flicker Noise 2 0 0.10 0.32 1.00 3.16 10.00 31.62 100.00 Normalized Frequency, f/fc Figure (9). The noise currents in the drain-source port of the MOSFETs modeled in Figures (7b) and (8b). Corner frequency fc is defined such that at this frequency, the flicker and drain current noise components are identical. The drain and flicker noise currents are presumed uncorrelated. The normalization factor for the mean square noise current is (4kTγgmΔf). A simplification, as it were, is to refer the net drain-source noise current, jnds(t), in the noise macromodel of Figure (10a), to the gate circuit as a voltage, ends(t), which we delineate in Figure (10b). Since gate-source signal voltages are multiplied by forward transconductance to produce drain signal currents, we argue that the mean square value of this reflected noise voltage, ends(t), is Electronic Circuit Noise, Part I_R1 - 21 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma 4kT 1 c 2 j (t) f e 2 (t) nds Δf . nds 2 g g m m f (61) An important design-oriented motivation underlies the foregoing transformation. In particular, and within the constraint of uncorrelated noise sources, the mean square value of the reflected noise, as per (61), simply adds to the mean square value of the gate noise, which (50) defines analytically. As a result, the two gate circuit noise voltages in Figure (10b) can be replaced by the equivalent single noise voltage, en(t), shown in Figure (10c), with the understanding that the mean square value of en(t) is Noiseless IdQ rg * eng(t) jnds (t) ind (t) jnf (t) Vds Vgs rg IdQ Noiseless eng(t) * ends(t) * (a). Vds Vgs rg * en(t) IdQ Noiseless (b). Vds Vgs (c). Figure (10). (a). The MOSFET noise macromodel of Figure (7c). (b). Noise currents in the drain-source circuit of the noise model in (a) referred to the gate circuit as a noise voltage, ends(t). The two gate circuit noise voltages are uncorrelated. (c). Superposition of the two noise voltage sources in the gate circuit of the model in (b) into an equivalent, single noise voltage, en(t), in the gate circuit. Electronic Circuit Noise, Part I_R1 - 22 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma e 2 (t) e 2 (t) e 2 (t) n nds ng f f 1 c 1 c 5 (62) f f 4kT r Δf 4kT Δf , g g g m m where we have exploited (51). Aside from simplifying the noise macromodel of Figure (10a) to the single noise generator architecture depicted in Figure (10c), (62) serves the important purpose of effectively defining the maximum tolerable input noise, or noise floor, of a simple common source amplifier. In such an amplifier, the signal to be processed is a voltage placed in series with the gate terminal of the transistor, which is to say that this signal voltage appears in series with the equivalent input noise voltage, vn(t). It follows that if the amplifier is to detect reliably the signal voltage, the applied signal must overcome, in addition to its own internally generated noise, the RMS value, Enrms, of the equivalent input noise voltage. At a minimum, therefore, this RMS noise voltage must be smaller (and preferably, substantially smaller) than the RMS value, Vsrms, of the applied signal; namely, f 1 c 5 f E e 2 (t) 4kT (63) Δf Vsrms . nrms n g m The result at hand implies the need for large transconductance and small flicker corner frequency if the noise inherently produced by a MOSFET is precluded from significantly contaminating the input signal identified for appropriate processing. 3.3.3.4. Another Source of MOSFET Noise In addition to the drain noise we have elucidated in Section (3.3.3.1), thermal noise associated with the substrate resistance can superimpose with the drain noise current, jnd(t), in Figure (10a)[7]. The mean square value of this substrate thermal noise, which is oft referred to as epitaxial noise, is 2 4kTR g 2 sub mb Δf , (64) j (t) sub 2 1 2πf RsubC gb where Rsub is the average value of the bulk substrate spreading resistance, Cgb represents gatebulk capacitance, and gmb designates the bulk transconductance (partial derivative of drain signal current with respect to bulk-source signal voltage). For most, but assuredly not all, analog MOSFET applications, substrate noise is likely to be inconsequential for three reasons. First, the bulk transconductance, gmb, is ideally zero and likely to at least be a very small transconductance, especially if the MOSFET is fabricated in a monolithic process that features very thin gate oxides[9]-[10]. Second, substrate resistance Rsub, can be reduced by prudent layout strategies entailing the liberal use of substrate electrical contacts that are returned to circuit ground. Third, at high frequencies, where capacitance Cgb effectively bypasses the substrate resistance, the Electronic Circuit Noise, Part I_R1 - 23 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma denominator on the right hand side of (68) becomes large, thereby rendering the mean square value of the substrate noise current inconsequential. 4.0. TWO PORT NOISE ANALYSIS The majority of analog MOSFET amplifiers behave electrically as linear two port networks. As such, their I/O gains are independent of signal amplitudes and noise levels, thereby allowing the convenience of referring all network noise sources to any system port and in particular, to the system input port. When we speak of an “input port referred noise,” or more simply, “input referred noise,” we are talking about one or perhaps two effective noise sources. The mean square values of these effective noise generators are judiciously chosen so that when they activate the input port of the noiseless equivalent of the network we are studying, they deliver the same net output noise as observed in the original network, which is doubtlessly permeated by numerous internal sources of electrical noise. Thus, the action of referring the net noise level at the output port of a network to the network input port serves to simplify circuit analysis in that potentially many noise sources intrinsic to the two port network are supplanted by a mere one or two input port noise generators. Additionally, once we have succeeded in referring all network noise sources to the input port, we can straightforwardly compare the strength of an applied signal to the noise source (or sources). Aside from ascertaining the noise floor, the careful examination of the nature of the effective input noise sources is likely to forge meaningful design strategies aimed toward minimizing the deleterious effects of noise. Such minimization is tantamount to reducing the noise floor. Figure (11) abstracts the foregoing declarations. In Figure(11a), we show a linear, noisy, two-port network whose applied input signal is represented by a Thévenin equivalent circuit comprised of voltage source Vs, and impedance Zs. Additionally, parasitic signal noise in the form of random voltage ens(t), whose mean square value is e 2 (t) , superimposes with, and therens fore blurs, the time deterministic signal, Vs. The output port of the network is terminated in the load impedance, Zl. In response to the input signal, the indicated signal noise, noise sources generated within the linear two-port network, and any noise generated by the load termination, the net input port voltage consists of signal voltage Vi superimposed with noise voltage eni(t). Analogously, the net output port voltage response superimposes a signal component, Vo, and a noise voltage, eno(t). In Figure (11b), we examine only the noise responses of the system diagrammed in Figure (11a) by setting the input time deterministic signal, Vs, to zero. Additionally, we set all noise sources within the two-port configuration and any noise implicit to the load termination to zero. As a result, the indicated two-port is the noiseless counterpart of the original two-port network in Figure (11a) and, of course, Zl is the noiseless equivalent of the original load termination. In place of these zeroed noise sources, we advance two noise sources, an equivalent input noise voltage, en(t), and an equivalent input noise current, in(t), at the input port, as shown in the diagram. These “equivalent” noise sources are selected so that the output noise response, vno(t), remains identical to the output port noise voltage observed in the original system of Figure (11a). The determination of the equivalent input noise voltage and current first requires the computation or measurement of the mean square value, e 2 (t) , of the net output noise voltage, no eno(t). Such a computation must judiciously consider the noise associated with the applied signal Electronic Circuit Noise, Part I_R1 - 24 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma Vs Zin * ens(t) Zl Potentially Noisy Noisy, Linear Two-Port Network source, the noise sources implicit to the two-port network and the terminating load impedance, any frequency dependencies of all of these noise sources, and finally, correlation factors evidenced between any two noise sources. Then, if the input port -to- output port voltage gain is Av(j2πf), which, because of linearity, can be deduced from signal considerations alone as Vi + eni (t) Vo + eno (t) Zs ens(t) * eno (t) Noiseless jn (t) Zin Zs * eni (t) Zl en (t) Noiseless, Linear Two-Port Network (a). (b). Figure (11). (a). A noisy, linear two-port network whose applied input excitation is the superposition of a time deterministic signal, Vs, and a noise voltage component, ens(t). (b). The noiseless equivalent of the noisy two-port network in (a). The equivalent noise voltage and noise current sources, en(t) and jn(t), respectively, are selected such that in conjunction with the input signal noise, ens(t), the total output port noise voltage, eno(t), is identical to the output noise response observed in (a). A (j2πf) V o , V i v (65) the mean square value (or square of the corresponding RMS value) of the input port noise voltage, vni(t), corresponding to output noise voltage, vno(t), is e 2 (t) ni e 2 (t) no A (j2πf) 2 . (66) v Electronic Circuit Noise, Part I_R1 - 25 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma But we observe in Figure (11b) that the input port noise voltage, eni(t) is identically equal to noise voltage en(t) under the conditions of zero signal noise and zero source impedance. Hence, we arrive at the mean square value of the equivalent input noise voltage by computing the mean square value, e 2 (t) , of the equivalent input noise voltage, en(t), in accordance with n e 2 (t) e 2 (t) v (t)0 n ni ns Z 0 e 2 (t) no 2 A (j2πf) v s . (67) v (t) 0 ns Z 0 s We also observe in Figure (11b) that if the source impedance, Zs, is infinitely large, the input port noise voltage is necessarily produced exclusively by the flow of noise current jn(t) into the input port of the network. Thus, with j 2 (t) representing the mean square value of the equivan lent input noise current, jn(t), e 2 (t) ni Z (j2πf) in Z 2 2 j (t) , n (68) s whence, e 2 (t) ni j 2 (t) n Z s Z (j2πf) e 2 (t) no 2 2 Z (j2πf) in in A (j2πf) v . 2 (69) Z s Equations (68) and (69) establish the mathematical basis for identifying the mean square values (or the square of the RMS values) of the equivalent input noise voltage and the equivalent input noise current that act to emulate the original output port noise response of the considered twoport network. 4.1. NOISE FLOOR In Figure (11b), let the RMS values of signal source noise voltage ens(t), equivalent input noise voltage en(t), equivalent input noise current jn(t), and input port noise voltage eni(t), be denoted by Ensrms, Enrms, Jnrms, and Enirms, respectively. By superposition, Z Z Z Z in in E E (70) E in s J , nirms Z Z nsrms Z Z nrms Z Z nrms in s in s in s where we have elected to add all of the three pertinent voltage components because, as we have already observed, phase angles between truly random voltage and current components are indeterminate. As it materializes, it is more convenient to couch (70) into the form, E nirms J nsrms J nrms Y E s nrms Y Y in , (71) s where Yin is the admittance corresponding to input impedance Zin, Ys is the admittance of source impedance Zs, and Electronic Circuit Noise, Part I_R1 - 26 - July 2011 Technical Report #02-0511_R1 J nsrms E University of Southern California Viterbi School of Engineering Choma nsrms Y E s nsrms Z s (72) is the RMS Norton current equivalent of RMS signal noise voltage Ensrms. The signal noise current, jns(t), is doubtlessly not correlated with either jn(t) or en(t). But noise current jn(t) and noise voltage en(t) are correlated because from (67) and (69), the mean square values of both of these noise generators are functionally dependent on the mean square output noise voltage and the network voltage gain. Thus, the mean square noise voltage developed across the input port of the linear network is J 2 2 e (t) E ni nirms 2 2 J Y E nsrms nrms s nrms , 2 Y Y in s (73) where we have exploited the fact that the mean square values (or squares of RMS values) of only uncorrelated random variables directly superimpose with one another. It is understood that both Yin and Ys in (73) are complex functions of frequency; that is, Y Y (j2πf) s s (74) . Y Y (j2πf) in in It now follows that the mean square value of the net output noise of the linear two-port network before us is 2 2 J J Y E 2 2 nrms s nrms A (j2πf) E 2 A (j2πf) nsrms e 2 (t) E 2 . (75) no norms v nirms v 2 Y Y in s Despite the messiness of this relationship, the mean square output noise voltage, say e 2 (t) , nos attributed solely to the noise associated with the applied input signal is clearly revealed as e 2 (t) nos A (j2πf) 2 v J Y Y in 2 nsrms , 2 (76) s which implies that the corresponding mean square input port noise, E 2 , due exclusively to nis noise contamination implicit to the signal source is e 2 (t) nis e 2 (t) nos A (j2πf) v 2 J2 nsrms Y Y in 2 . (77) s In Figure (11a), we see that the signal delivered to the input port of the linear two-port network is Z I in V s (78) V , i Z Z s Y Y s in s in where Electronic Circuit Noise, Part I_R1 - 27 - July 2011 Technical Report #02-0511_R1 I s University of Southern California Viterbi School of Engineering V s YV s s Z s Choma (79) is the Norton equivalent signal current corresponding to the applied signal voltage, Vs and its intrinsic Thévenin impedance, Zs. In order for the two-port network to detect signal voltage Vi faithfully and ultimately process it in accordance with the design targets stipulated for the twoport system, Vi must rise above the RMS noise voltage, Enirms, established across the network input port. Thus, we require Vi > Enirms and from (75), (76), and (83), this mandate is seen as implying V E s nsrms Z J s nrms E nrms . (80) The minimum detectable signal, Vsmin, which defines the noise floor of the two-port network therefore evolves as V smin E nsrms Z J s nrms E nrms . (81) Thus, in addition to the self-evident observation that the input signal must overcome its own noise level, whose RMS amplitude is Ensrms, it must also overcome the deleterious effects of the equivalent input noise voltage and current, whose combined impact on the input port is measured as the voltage sum, (Zs Jnrms + Enrms ). 4.2. NOISE FACTOR Recalling (7) -through- (9), the noise factor, F, of the two-port network displayed in Figure (11a) is V2 E2 i irms norms F 2 SNR V E2 o orms nirms SNR E2 1 A (j2πf) v 2 norms , E2 (82) nirms where we have applied (65). Before proceeding further, there may be an advantage, from the perspective of simply expediting recollection, to expressing noise factor F in (82) in terms of simple engineering terminology. In particular, a careful investigation of the right hand side of (82) allows us to proffer, Net Mean Square Output Noise Voltage (Current) Due To All Noise Sources F . (83) Net Mean Square Output Noise Voltage (Current) Due Only To Source Noise Appealing to (75) and (77), we arrive at the relatively simple analytical expression, F SNR i SNR J Ys Enrms 1 nrms o J2 2 . (84) nsrms The result at hand confirms earlier contentions to the effect that the noise factor of a network exceeds one but approaches one when the noise implicit to the network undergoing investigation is negligible in comparison to the noise implicit to the signal source. This implicit network noise is determined by equivalent input noise current jn(t) and equivalent input noise voltage en(t), whose RMS values are respectively Jnrms and Enrms, as deployed in (84). Using (72), we can also cast (84) directly in terms of the noise voltage that accompanies the input signal. Specifically, Electronic Circuit Noise, Part I_R1 - 28 - July 2011 Technical Report #02-0511_R1 F 4.3. SNR i SNR o University of Southern California Viterbi School of Engineering J Ys Enrms 1 nrms Y (j2πf) 2 s 2 2 E nsrms 1 Z (j2πf) s 2 Choma J nrms Ys Enrms 2 E nsrms 2 . (85) NOISE FACTOR OPTIMIZATION In an attempt to formulate design strategies that ensure noise factor minimization in linear active networks, let us begin by decomposing the RMS value, Jnrms, of the equivalent input noise current into an uncorrelated RMS component, Jurms, and a component, say Jncrms, that remains correlated to the RMS equivalent input noise voltage, Enrms. In an actual application, current component Jurms may be null, or current Jncrms may be null, independent of the status of current Jurms. We write, J2 nrms Jurms J ncrms 2 Jurms Ync Enrms 2 , (86) where Ync is, in general, a frequency dependent complex admittance possessed of a real, or conductive component, Gnc, and an imaginary, or susceptive, component, Bnc, such that Y Y ( j2πf) G jB , (87) nc nc nc nc If we decompose the source admittance, Ys, into its conductive and susceptive components, Gs and Bs, respectively, such that Y Y ( j2πf) G jB , (88) s s s s (86) -through- (88) inserted into (84) delivers F 1 J 2 urms Gnc Gs j Bnc Bs J 2 nsrms 2 E 2 nrms . (89) In order to render (89) manageable in design environments, we shall introduce an equivalent conductance, Gu, such that the uncorrelated component of the equivalent input noise current is viewed as simple thermal noise generated by said conductance, Gu. This is to say that J 2 4kTGu Δf . urms (90) Similarly for the equivalent input noise voltage, E 2 4kTRn Δf . nrms (91) Finally, we shall presume that the noise current produced by the applied signal source can be attributed exclusively to source conductance, Gs; that is,4 J2 nsrms 4kTGs Δf . (92) Accordingly, (89) becomes 4 Attributing the signal noise to source conductance is a common engineering tack adopted in the characterization and assessment of the noise properties of a standalone amplifier or other type of linear two-port network. But in an actual multistage system, the signal noise is effectively the output port noise of the preceding stage, which can be significantly larger than mere thermal noise attributed to the output resistance of the preceding stage. Of course, the Thévenin voltage and output resistance of the preceding stage serves as the signal source for the network under investigation. Electronic Circuit Noise, Part I_R1 - 29 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering 2 2 G G G B B R u nc s nc s n F 1 . G Choma (93) s Since the susceptive part of a physically realizable admittance can be positive or negative, noise factor minimization requires that source susceptance Bs be equal to −Bnc. The resultant noise factor expression is infinitely large for both Gs = 0 and Gs = ∞, which implies that an optimum source conductance component, say Gsopt, exists. This optimal source conductance is straightforwardly deduced by setting to zero the first derivative of F with respect to Gs in (93), with Bs = −Bnc. The result is G G u (94) G G 1 G2 u , sopt nc nc R 2 R G n n nc for which the corresponding minimal (or optimized) noise factor, Fmin, is found to be G . u F 1 2R G G 1 2R G 1 1 min n nc sopt n nc 2 R G n nc (95) Design insights evolve straightforwardly from the foregoing results if we assume that the equivalent input noise current displays no correlation with its counterpart equivalent input noise voltage. In such an event, whose realism mandates a careful examination of the specific application under investigation, admittance Ync in (86) and (87) is null. Obviously, the conductive component, Gnc, and susceptive component, Bnc, of admittance Ync are necessarily zero. With Bnc = 0, noise figure minimization in (93) requires that the signal source admittance be a purely real conductance; that is, Bs = −Bnc = 0, whence Ys in (88) is simply Gs. With Gnc = 0, the optimum value, Gsopt, of signal source conductance Gs is, by (94), (90), and (91), G J J u urms nrms . (96) G sopt Y 0 R E E n nc nrms nrms In this expression, we have exploited the fact that if equivalent input noise current, jn(t) and equivalent input noise voltage en(t) are uncorrelated, the RMS value, Jurms, of the uncorrelated component of jn(t) is simply the RMS value, Jnrms, of jn(t). In short, the optimum impedance value of a signal source applied to a noisy, linear active network displaying no correlation betwixt its equivalent input noise current and voltage is simply the ratio of the RMS values of equivalent input noise voltage and equivalent input noise current. In the interest of completeness, we note that the minimum noise factor corresponding to (96) is, from 95), F min Y 0 1 2R G n sopt nc 1 2 G R . (97) u n It is enlightening to witness that (93) can be massaged into the algebraic form, 2 G 2 G G B B s FF , s sopt s nc min R n (98) which asserts that in the source admittance plane, the contours of constant noise factor are circles centered at (Gs, Bs) = (Gsopt, −Bnc), with radii, r(F), of Electronic Circuit Noise, Part I_R1 - 30 - July 2011 Technical Report #02-0511_R1 r(F) G University of Southern California Viterbi School of Engineering Choma s F F . min n R (99) These contours are displayed generically in Figure (12). Observe in this figure, as predicted, that the noise factor is minimized at its optimal value of Fmin when Gs = Gsopt and simultaneously, Bs = −Bnc. Unfortunately, Gsopt and Bnc are likely to be frequency dependent owing to flicker phenomena implicit to the considered linear active network and the likelihood that the network transfer function, which is pivotal to the determination of the equivalent input noise voltage and current, is a function of frequency. Accordingly, we anticipate encountering challenges if attempts are made to achieve broadband noise factor optimization. On the other hand, such optimization is more easily accomplished at a single frequency and perhaps even over a restricted passband. Bs F3 F2 F1 Gsopt 0 Bnc Gs Fmin Figure (12). The contours of constant noise figure for a linear two-port network. A signal source admittance, Ys, which equals (Gsopt − jBnc), delivers an optimal noise figure of Fmin. In radio frequency (RF) amplifier applications, stereotypically anemic signal strengths encourage conjugate matching of the signal source impedance to the amplifier driving point input impedance. This matching assures maximum signal power transfer from the signal source to the amplifier input port, thereby guarding against excessive signal power losses that derive from inefficient signal coupling. Unfortunately, the source admittance corresponding to maximum signal power transfer at the amplifier input port is rarely the source admittance deemed optimal for noise factor optimization[11]. Accordingly, a design compromise for high performance RF amplifiers is mandated to preclude both excessively large noise factor and inappropriately large signal power losses between signal source and input port. Electronic Circuit Noise, Part I_R1 - 31 - July 2011 Technical Report #02-0511_R1 5.0. University of Southern California Viterbi School of Engineering Choma NOISE FACTOR OF A CASCADE Most amplifiers are multistage units, as is underscored pictorially by the symbolic three-stage cascade of Figure (13a). Since several noise sources are likely to appear in each stage, the noise analysis of the entire cascade is sufficiently cumbersome to impede the formulation of relevant design insights. For this reason, the development of an efficient and mathematically tractable method of noise characterization in complex linear network architectures is an enviable undertaking. To this end, we shall derive an expression for the overall noise factor, say F, in terms of the individual stage noise factors, Fi, and the associated stage voltage gains, Avi. In the subject figure, Eni denotes the RMS value of the output noise generated exclusively by the noise sources internal to the ith amplifier stage. Eno1 Eno2 Ens Eno Stage #1 Stage #2 Stage #3 {En1 , Av1 , F1 } {En2 , Av2 , F2 } {En3 , Av3 , F3 } (a). Eno1 Eno2 Stage #1 Ens * Stage #2 Ens {En1 , Av1 , F1 } Stage #3 Ens * * {En2 , Av2 , F2 } Eno3 {En3 , Av3 , F3 } (b). Figure (13). (a). Cascade of three linear amplifier stages. In general, the ith stage is characterized by an RMS output noise voltage due to internal noise sources of Eni, an I/O port voltage gain of Avi, and a noise figure of Fi. (b). The noise characterization of each of the three amplifier stages in (a). The reference input noise to each stage is identical and has an RMS voltage of Ens, which is the same as the RMS noise voltage applied to the input port of the cascade in (a). In the cascade of Figure (13a), a test input noise source of RMS value Ens is applied to the cascade input port. In order to effect a stage noise characterization that is consistent with the noise properties deduced for the entire network cascade, we evaluate the noise performance of each stage by applying this same noise source to the input ports of the individual stages. This Electronic Circuit Noise, Part I_R1 - 32 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma engineering methodology is abstracted in Figure (13b). Then, using (83) as a basis for an analytical noise factor expression and noting that En1 is the RMS value of the output noise attributed solely to the noise internal to the first stage, we have5 A 2 2 2 E ns n1 , 2 2 A E v1 ns v1 F 1 E (100) from which we deduce E2 n1 F1 1 Av1 2 E2 . (101) ns It is hardly surprising that the mean square internal noise generated in the first stage is zero if F1 = 1. A similar analytical tack yields for the mean square noise voltages incurred by internal noise sources in the second and third stages, E2 F2 1 Av2 2 E2 F3 1 Av3 2 n2 and n3 E2 , (102) E2 . (103) ns ns In order to evaluate the noise factor of the entire three-stage system, we need first to find the net mean square output noise, E 2 , produced at the network output port. In arriving at no this metric, we note several important points in Figure (13a). First, the test noise applied at the cascade input port is amplified by all three stages. However, the output noise caused by stage #1 internal noise sources is amplified by only the second and third stages. Similarly, the output noise incurred by internal noise sources within the second stage is amplified by only the third stage, while the internal third stage noise is not amplified. Accordingly, we stipulate E2 no A A A v1 v2 v3 2 E2 A A ns 2 v2 v3 E2 A n1 2 v3 E2 E2 . n2 (104) n3 Since the output mean square noise generated by only the signal source noise applied to the input port is the first term on the right hand side of the this equation, we deduce an overall noise factor, F, of E2 A A A E2 A A ns v2 v3 2 E2 A n1 F 2 2 2 2 A A A E A A A E v1 v2 v3 ns v1 v2 v3 ns E2 E2 E2 n1 n2 n3 1 . 2 2 2 2 2 2 A E A A E A A A E v1 ns v1 v2 ns v1 v2 v3 ns no v1 v2 v3 2 v3 2 E2 E2 n2 n3 (105) Using (101)-(103), this relationship can be written in the streamlined form 5 Metric Ens is the RMS voltage of an independent test noise source applied to the the input port of the entire cascade in Figure (13a), which is also the noise source activating the individual test stage structures in Figure (13b). This independent noise source can be presumed to be uncorrelated to the internal amplifier noise that generates RMS output voltage components of Eni in the ith stage. Electronic Circuit Noise, Part I_R1 - 33 - July 2011 Technical Report #02-0511_R1 F F 1 University of Southern California Viterbi School of Engineering F 1 2 A v1 2 F 1 3 A A 2 Choma , (106) v1 v2 which the literature archives as Friis’ formula[12]. Obviously, (106) can be extended to any number of stages; namely, F 1 F 1 F 1 F 1 3 4 n F F 2 , (107) 1 2 2 2 2 A A A A A A A A A A v1 v1 v2 v1 v2 v3 v1 v2 v3 v(n-1) where, of course, n is the number of stages. Equation (106), or its generic form in (107) is interesting from several engineering perspectives. First, it argues that the noise factor of a cascade of linear network stages can be determined unambiguously in terms of only the individual stage noise factors and the squared voltage gain magnitude (or power gain) of these individual stages. Second and perhaps most importantly, the Friis’ formula shows that if the squared voltage gain magnitude of the first stage in the cascade is sufficiently large, the overall noise factor is largely determined by only the noise factor of the high-gain first stage. This means, within reason, that a low noise design essentially boils down to the low noise design of the first stage. To the latter end, note in (107) that the noise factors of stages 2 -through- n contribute to the individual terms on the right hand side of (107) as a noise factor less one, whereas the contribution of the first stage to overall noise factor is the full noise factor of that stage. Finally, we observe that since the squared magnitude of the first stage voltage gain appears in the denominator of every term beyond the first term on the right hand side of (107), a first stage gain magnitude of less than one over the noise equivalent passband is inappropriate. Thus, for example, a source follower as the first stage of an intended low noise amplifier design is hardly prudent. 6.0. REFERENCES [1]. J. B. Johnson, “Thermal Agitation of Electricity in Conductors,” Physics Rev., vol. 32, pp. 97109, July 1928. [2]. H. Nyquist, “Thermal Agitation of Electric Charge in Conductors,” Physics Rev., vol. 32, pp. 110-113, July 1928. [3]. C. D. Motchenbacher and F. C. Fitchen, Low-Noise Electronic Design. New York: John Wiley and Sons, Inc., 1973, pp. 172-175. [4]. D. A. Johns and K. Martin, Analog Integrated Circuit Design. New York: John Wiley and Sons, Inc., 1997, chap. 4. [5]. Y. W. Lee, Statistical Theory of Communication. New York: John Wiley and Sons, Inc., 1963, pp. 9-45. [6]. A. van der Ziel, “Thermal Noise in Field Effect Transistors,” Proc. IEEE, pp. 1801-1812, Aug. 1962. [7]. T. H. Lee, The Design of CMOS Radio-Frequency Integrated Circuits. United Kingdom: Cambridge University Press, 2004, chap. 11. [8]. K. R. Laker and W. M. C. Sansen, Design of Analog Integrated Circuits and Systems. New York: McGraw-Hill, Inc., 1994, pp. 79-85, 161. [9]. J. Choma, “The Metal-Oxide-Silicon Field Effect Transistor,” University of Southern California, EE 536a Course Notes #1, July 2008, available at http://www.jcatsc.com/. Electronic Circuit Noise, Part I_R1 - 34 - July 2011 Technical Report #02-0511_R1 University of Southern California Viterbi School of Engineering Choma [10]. J. Choma, “Circuit Level Models and Basic Applications of MOS Technology Transistors, University of Southern California, EE 536a Lecture Aid #1, 2010-2011, available at http://www.jcatsc.com/. [11]. D. Shaeffer and T. Lee, “A 1.5V, 1.5 GHz CMOS Low Noise Amplifier,” IEEE J. Solid-State Circuits, May 1997. [12]. J.D.Kraus, Radio Astronomy. New York: McGraw-Hill Book Company, 1966 Electronic Circuit Noise, Part I_R1 - 35 - July 2011