Parts 1–2 introduced noise sources and their spectral densities. This lecture shows how to build
circuit-level noise models of real semiconductor devices: first the diode, then the bipolar junction
transistor (BJT). Both are non-linear, so the analysis starts from a small-signal (linearised) model and attaches the
physical noise mechanisms to it. The lecture ends with the subtle part — why a BJT’s input-referred noise needs
a voltage and a current source, and why they cannot be treated as independent.
A diode is a non-linear device described by the Shockley equation:
Shockley diode equationID = IS(eVD/ηVT − 1), VT = kBT/q ≈ 25.7 mV at room temperature
with IS the reverse saturation current (order of 10−10 A), η the ideality factor (typically 1–2)
and VT the thermal voltage. The equation assumes steady state — it says nothing about how the diode responds to
time-varying signals.
We study small signals because they are the ones most susceptible to noise. For a small signal the diode is usually
forward biased, and in the small region around the operating point Q the curve can be replaced by its tangent:
Slope at the Q-point
dID/dVD |Q = (IDQ + IS)/(ηVT) ≈ IDQ/(ηVT)
(since IS ≈ 10−10 A ≪ IDQ)
This slope has units of conductance, so its inverse is a resistance — the dynamic resistancerd = ηVT/IDQ. To capture time-varying behaviour the junction capacitance Cj and the
diffusion capacitance Cd are placed in parallel with rd.
2. Diode noise sources
A diode has three main noise sources, all attached to the small-signal model:
Shot noise — the PN junction is a potential barrier that carriers must cross, so the diode current has shot noise, modelled as a parallel current source across the junction (hence across rd):
īns2 = 2qIDQΔf, with IDQ the quiescent current.
Flicker noise — being a semiconductor device it has 1/f noise, also a parallel current source:
īnF2 = KF IDQα ln(fH/fL), where KF and α are device-dependent constants.
Thermal noise — the contacts and silicon substrate have a real resistance Rs, modelled as a resistor in series with rd that generates a series voltage noise v̄nRs2 = 4kBT RsΔf.
Fig. 1 — Small-signal noise model of a diode (redrawn from the lecture slides)
Why doesn’t rd make thermal noise? The dynamic resistance is not a real resistor — it is a
behavioural quantity added to the model to resemble the diode’s I–V relationship. Only physical resistances (like the bulk resistance Rs) generate thermal noise.
3. Total diode noise voltage
The three sources sit in different branches, so they cannot simply be added. Rs’s thermal noise is a series voltage, whereas shot and flicker noise are
parallel currents. The fix is to convert the parallel currents into an equivalent series (Thévenin) voltage across rd, after which KVL applies and, with the sources uncorrelated, the mean-square values add:
Total noise voltage of a diode (with Δf = fH − fL)v̄n,tot2 = 4kBTΔf Rs + [ 2qIDQΔf + KF IDQα ln(fH/fL) ] rd2
Substituting IDQ = ηVT/rd and VT = kBT/q collapses the shot-noise term and gives the compact form from the lecture:
For simplicity the capacitances are ignored here (open circuit), which is reasonable for frequencies up to the MHz range where their roll-off has not yet begun.
4. BJT small-signal model and current gain
There are several small-signal models for a BJT. The lecture groups them into physics-based models (the hybrid-π model, with simplified, low-frequency and high-frequency variants, and the T-model) and
two-port / black-box models (h-parameters, built from measured terminal behaviour). The analysis uses the high-frequency hybrid-π model with
gm, rπ, ro, Cπ and Cμ (Fig. 2).
Linearising IC ≈ ISeVBE/VT about the operating point gives the transconductance gm = ICQ/VT (collector shorted so ro does not matter), and the DC current gain follows directly:
DC current gain βDC = IC/IB = gm rπ
With the collector shorted for AC, the base sees rπ in parallel with Cπ and Cμ, so the AC current gain is a low-pass response:
AC current gain and unity-gain frequency
βAC = βDC / (1 + jωrπ(Cπ+Cμ)),
|βAC|2 = 1 / [ 1/βDC2 + (ω/ωT)2 ], ωT = gm/(Cπ+Cμ)
This resembles an RC low-pass filter with corner f−3 = 1/[2πrπ(Cπ+Cμ)]. As ω approaches ωT the current gain falls to 1: beyond that the BJT no longer works as an amplifier, so ωT is the maximum useful operating frequency. Because gm = ICQ/VT, the unity-gain frequency is not a fixed property of the device — it changes with the quiescent collector current.
5. The six noise sources of a BJT
A BJT has six main noise sources: three thermal, two shot and one flicker.
Thermal noise from the bulk resistances of the emitter, base and collector regions (RE, RB, RC), added in series with the corresponding terminals.
RE and RC are generally ignored: the emitter is heavily doped (minimal resistance), and the collector, although lightly doped, has a wide cross-section and a well-doped substrate to carry signals outward.
RBcannot be ignored: the base is a very narrow, only moderately doped region, so its resistance is comparatively high, and noise generated there is amplified by β.
Shot noise: there are two current paths, and each crosses a potential barrier. Forward-biasing the base–emitter junction sends electrons from the highly doped emitter into the thin base; almost all diffuse across and are swept into the collector, so the collector current is a stream of carriers crossing a barrier — collector shot noiseīnC2 = 2qICQΔf, in parallel with the C–E junction. The small difference between IE and IC is the base current IB (a few charges recombine in the base); it crosses the B–E junction and gives base shot noiseīnB2 = 2qIBQΔf, in parallel with the B–E junction.
Flicker noise: as with any semiconductor, trap states give 1/f noise, modelled as a current source in parallel with the B–E junction: īnF2 = K IBQα ln(fH/fL).
Fig. 2 — Hybrid-π model of a BJT with its noise sources (redrawn from the lecture slides). RE and RC are neglected.
As with the diode, the dynamic resistances rπ and ro are not real resistors and add no noise.
6. Referring BJT noise to the input
A BJT is a two-port, so its equivalent input-referred noise model needs a series voltage source and a parallel current source across the input port. Everything is transferred to the input side:
base thermal noise is already in series at the input; base shot and flicker noise are already across the input; only the collector shot noise is still at the output.
To refer the collector shot noise to the input as a current, divide by the squared magnitude of the AC current gain; to refer it as a voltage, divide by the squared transconductance (an AC quantity):
Collector shot noise referred to the inputīnc−ir2 = īnC2/|βAC|2 = 2qΔf IBQ [1 + (ωrπ(Cπ+Cμ))2] / (gm rπ) v̄nc−ir2 = īnC2/gm2 = 2qΔf IBQ rπ/gm
Adding up the pieces, the complete input-referred noise model of the BJT is:
The additions above are valid only when the sources have zero mean and are uncorrelated, and when KVL/KCL permit the addition. In reality the collector and base noise are somewhat correlated, but for this course they are treated as uncorrelated.
7. Why both a voltage and a current source are needed
This is the conceptual heart of the lecture. The input-referred collector shot noise appears twice in the model — as īnc−ir and as v̄nc−ir — but they are the same noise source, so they are not independent:
when the total noise at the output is calculated, the contributions of these two terms cannot simply be added as squared quantities. Their correlation matters.
The calculation uses superposition for each source with the signal source replaced by its source resistance RS (at low frequency, so Cπ and Cμ drop out). The result for the output noise current of a common-emitter stage is:
The last term is the cross-correlation between the voltage and current forms of the collector shot noise. Two observations follow:
Ideal source (RS = 0): the expression collapses to īout2 ≈ gm2(v̄nB2 + v̄nc−ir2), and it is the base thermal noise and the voltage form of the collector shot noise that dominate — the base shot and flicker noise vanish because RB is small.
Non-zero source resistance: the current form of the collector shot noise now matters too, because it flows through RS. If only the current-source part were propagated to the output (or only the voltage part), the result would not equal the true collector noise current īnC; with both parts and the correlation term, it does.
The takeaway: this is exactly why two-port theory demands both an equivalent voltage and an equivalent current noise source at the input — a single source cannot give correct answers for every source impedance a user might connect. It also shows how the dominant noise source of a BJT amplifier depends on the source impedance.
8. Exercise and take-homes
Exercise 1 (from the lecture). A BJT amplifier has β = 50, gm = 38 mS and rπ = 1.33 kΩ (1 mA bias), with all capacitors noiseless and short-circuited for AC. Measurements on the amplifier give a noise figure of 3 dB with a source resistance of 50 Ω and 9 dB with 500 Ω; the effective bandwidth is 100 kHz. Find the transistor’s input-referred voltage and current noise spectral densities (v̄n2/Hz and īn2/Hz), assuming both are white.
Important take-homes from the lecture:
Noise is fundamentally an AC quantity and a small signal in itself; begin with small-signal models for analysis.
Dynamic resistances in small-signal models do not generate noise — they represent behaviour, not physical resistors. Bulk resistances and terminal resistances are physical and must be considered.
The capacitances were mostly ignored during these derivations; think about how including them would change the noise characteristics (effective bandwidth).
A diode has thermal, shot and flicker noise. A BJT has several intrinsic noise sources, of which four are important (base thermal, base shot, collector shot, flicker).
You must be able to calculate the AC current gain βAC and explain why it is needed; refer noise sources to the input so a generic noise model results; and understand why the 2-port model needs both a voltage and a current source at the input.
Explain how source impedance changes which noise sources (voltage or current) dominate; use superposition to propagate individual sources, and recognise when input-referred sources are not independent.
Derivations are not examinable in detail — but the concepts used (and why each simplification is valid) are.
Click text to edit it. Drag a photo's ≡ handle to reorder it within its row. Nothing saves until you hit Export — that downloads the edited file for you to drop back into your Drive folder.