Electronics Guide

Statistical Characterization

Since noise cannot be described by a specific time-domain waveform, statistical measures characterize its properties:

• Mean value: The average voltage or current over time; for most noise sources, this is zero
• Variance: The average of the squared deviation from the mean, representing noise power
• Root mean square (RMS): The square root of variance, providing a measure of typical noise amplitude
• Probability distribution: Describes the likelihood of various amplitude values; thermal and shot noise follow Gaussian distributions
• Power spectral density: Describes how noise power is distributed across frequency

The Gaussian (normal) distribution applies to most fundamental noise sources because they arise from the sum of many small independent contributions. The central limit theorem ensures that such sums approach a Gaussian distribution regardless of the distribution of individual contributions.

Gaussian noise has no finite peak value. Designers therefore work with a crest factor: the instantaneous amplitude stays within ±3σ of the mean about 99.7 percent of the time, so a peak-to-peak figure of 6σ is the usual engineering convention for converting an RMS noise figure into the excursion a circuit must tolerate.

Noise Power and Addition

When multiple independent noise sources contribute to a signal, their powers add rather than their amplitudes. For two uncorrelated noise voltages with RMS values V₁ and V₂:

V_total = √(V₁² + V₂²)

This root-sum-square addition reflects the fact that independent noise sources are uncorrelated. At any instant they may add constructively or destructively, but on average their powers combine. Adding RMS noise voltages arithmetically is a common and serious error: it overstates the total. The rule also has a useful consequence for design, because a noise source that is small compared with the dominant source has minimal effect on the total.

For example, if one noise source contributes 10 µV RMS and another contributes 3 µV RMS, the total noise is:

√(10² + 3²) = √109 ≈ 10.44 µV RMS

The smaller source increases the total by only about 4.4 percent, despite being 30 percent of the larger source in amplitude. Correlated sources are the exception: when two contributions share a common origin, a correlation coefficient enters the sum, and fully correlated voltages do add linearly.

Signal-to-Noise Ratio

Signal-to-noise ratio (SNR) quantifies how much a signal exceeds the noise floor. Expressed in decibels:

SNR (dB) = 20 log₁₀(V_signal / V_noise) = 10 log₁₀(P_signal / P_noise)

The factor of 20 applies to voltage or current ratios; the factor of 10 applies to power ratios. The two forms agree only when the signal and the noise are measured across the same impedance. An SNR of 60 dB means the signal power is one million times the noise power, or the signal voltage is one thousand times the noise voltage.

The minimum detectable signal depends on the required SNR, which varies by application. Voice communications may tolerate an SNR near 20 dB, while precision measurements might require 80 dB or more.

Physical Origin

At any finite temperature, electrons in a conductor possess thermal energy that causes them to move randomly. This random motion creates small fluctuating currents even in the absence of any applied voltage. The electrons scatter from the lattice and from one another, producing a continuous random motion that appears as a fluctuating voltage across the conductor's terminals.

Albert Einstein's early-twentieth-century work on Brownian motion and statistical fluctuations laid the groundwork for understanding such effects. John B. Johnson measured the phenomenon at Bell Telephone Laboratories and reported it in "Thermal Agitation of Electricity in Conductors," and Harry Nyquist supplied the theory in "Thermal Agitation of Electric Charge in Conductors," showing that the noise follows directly from thermodynamic equilibrium between the resistor and its electromagnetic environment. Both papers appeared in Physical Review, volume 32, in July 1928.

Because the result follows from equilibrium thermodynamics rather than from any device physics, it applies to every dissipative element: a resistor, the loss resistance of an inductor, the dielectric loss of a capacitor, and the radiation resistance of an antenna. A lossless reactance generates no thermal noise of its own.

Thermal Noise Equations

The open-circuit RMS noise voltage across a resistor is:

V_n = √(4kTRB)

Where:

• k = Boltzmann's constant, 1.380649 × 10⁻²³ joules per kelvin (an exact value since the 2019 redefinition of the SI base units)
• T = Absolute temperature in kelvins
• R = Resistance in ohms
• B = Noise bandwidth in hertz

That expression gives a total, integrated over the bandwidth B. The underlying quantity is a voltage spectral density, which carries no bandwidth at all:

e_n = √(4kTR)

The density is expressed in volts per root hertz. Keeping the two apart matters: a datasheet entry in nV/√Hz is a density and must be multiplied by the square root of the noise bandwidth before it can be compared with a total RMS figure.

At the standard reference temperature T₀ = 290 K adopted by IEEE for noise work, the density reduces to a convenient rule of thumb with R in ohms:

e_n ≈ 0.127 √R nV/√Hz

A 1 kΩ resistor at 290 K therefore produces about 4.0 nV/√Hz. At 300 K the coefficient rises to 0.129 and the same resistor produces about 4.07 nV/√Hz, so the temperature assumed behind any quoted figure is worth stating. Over a 10 kHz noise bandwidth at 290 K:

V_n = 4.0 nV/√Hz × √(10,000 Hz) = 400 nV RMS

Noise voltage grows only as the square root of resistance and of bandwidth. Cutting the noise voltage of a resistive source in half requires a fourfold reduction in either quantity.

White Noise Spectrum

Thermal noise has a flat power spectral density, meaning equal noise power per unit bandwidth at all frequencies. This characteristic is called white noise, by analogy with white light containing all colors equally.

The white spectrum extends from very low frequencies to extremely high ones, limited only by quantum effects that become significant when the photon energy hf approaches the thermal energy kT. The crossover frequency kT/h is about 6.0 THz at 290 K, far above the operating range of ordinary electronic circuits. Below that frequency the classical Nyquist result holds to excellent accuracy.

The flat spectrum simplifies noise calculations, because the total noise depends only on the noise bandwidth and not on the particular shape of the frequency response within that bandwidth.

Equivalent Noise Resistance

Any noise source can be characterized by its equivalent noise resistance, the resistance value that would produce the same thermal noise voltage at a stated reference temperature. For a source with voltage spectral density e_n:

R_eq = e_n² / (4kT)

This concept allows direct comparison of different noise sources and simplifies the combination of thermal noise with other noise types. An amplifier with 2 nV/√Hz of input voltage noise, referred to 290 K, has an equivalent noise resistance of about 250 Ω. Feeding that amplifier from a source resistance well below 250 Ω wastes its voltage-noise performance, because the amplifier rather than the source then sets the noise floor.

Thermal Noise in Circuit Analysis

For circuit analysis, a noisy resistor can be modeled in either of two equivalent ways:

• Thevenin model: A noiseless resistor in series with a voltage source of density e_n = √(4kTR)
• Norton model: A noiseless resistor in parallel with a current source of density i_n = √(4kT/R)

The choice between models depends on circuit topology. The Thevenin model is convenient when the resistor appears in series with the signal path; the Norton model suits parallel configurations. At 290 K a 1 kΩ resistor corresponds to 4.0 nV/√Hz in the Thevenin form and 4.0 pA/√Hz in the Norton form.

When resistors connect in series or parallel, their noise contributions follow from the equivalent resistance of the combination:

• Series resistors: R_total = R₁ + R₂, so V_n = √(4kT(R₁ + R₂)B)
• Parallel resistors: R_total = R₁R₂ / (R₁ + R₂), so V_n = √(4kTR_totalB)

A resistor shunted by a capacitor illustrates how the bandwidth term disappears. Integrating √(4kTR) over the response of the RC network gives a total mean-square noise voltage of kT/C, independent of R: the resistor that generates more noise density also narrows the bandwidth by the same factor. This kT/C result sets the floor for sampled-data circuits such as switched-capacitor filters and analog-to-digital converter sampling networks, where the only remedy is a larger sampling capacitor.

Physical Mechanism

Electrical current consists of discrete electrons, each carrying a charge of 1.602176634 × 10⁻¹⁹ coulombs, exact by SI definition. When these electrons cross a potential barrier, such as a p-n junction, they do so independently and at random times. That randomness creates fluctuations in the instantaneous current around its average value.

The key requirement for full shot noise is that carriers cross a barrier independently, so that their arrivals form a Poisson process. Ordinary metallic conduction does not satisfy it. In a plain resistor, carriers are not emitted over a barrier but drift through a medium in which collisions correlate their motion, so the resistor exhibits thermal noise and no shot noise, however large its direct current. Junctions in diodes, bipolar transistors, and photodiodes do satisfy the condition, and there shot noise appears.

Shot Noise Equation

The RMS shot noise current in a bandwidth B is:

i_n = √(2qI_DCB)

Where:

• q = Elementary charge, 1.602176634 × 10⁻¹⁹ coulombs
• I_DC = Direct current through the barrier, in amperes
• B = Noise bandwidth in hertz

The factor of 2 belongs to the single-sided spectral density used throughout engineering practice, in which all the noise power is folded onto positive frequencies. The corresponding density, again carrying no bandwidth, is:

i_n = √(2qI_DC)

For a 1 mA direct current the shot noise density is about 17.9 pA/√Hz, and over a 10 kHz noise bandwidth that gives about 1.79 nA RMS. Because the noise grows as the square root of current while the signal grows linearly, the signal-to-noise ratio of a shot-noise-limited detector improves as the square root of the current.

Shot Noise Spectrum

Like thermal noise, shot noise has a white power spectral density. The noise power per unit bandwidth is constant from low frequencies up to the point where carrier transit time across the barrier becomes comparable to the signal period. For most semiconductor junctions that limit lies in the gigahertz range or above.

Because both spectra are white and the two mechanisms are independent, shot noise and thermal noise add in quadrature:

i_total = √(i_shot² + i_thermal²)

Shot Noise in Different Devices

Shot noise behavior varies among semiconductor devices:

• Diodes: Full shot noise accompanies current crossing the junction, in both forward conduction and reverse leakage
• Bipolar transistors: Base current and collector current each generate full shot noise, and the base shot noise flowing in the source impedance is what makes bipolar noise performance depend so strongly on bias current
• Field-effect transistors: Channel current does not exhibit full shot noise, because carriers drift through the channel rather than crossing a barrier; the channel instead contributes thermal noise proportional to its transconductance, while gate leakage current does generate shot noise
• Photodiodes: Signal photocurrent and dark current both generate shot noise, frequently the dominant noise source in optical receivers

Shot Noise Suppression

Under certain conditions shot noise falls below the full Schottky value, because carriers are not truly independent:

• Space charge effects: In vacuum tubes and some semiconductor structures, accumulated charge correlates successive carrier arrivals
• Pauli exclusion: In degenerate conductors, occupied states are unavailable to other carriers, which smooths the current
• Quantum effects: In mesoscopic structures, quantum correlations reduce the fluctuations further

The Fano factor expresses the ratio of actual shot noise power to full Schottky noise. A Fano factor of 1 indicates full shot noise, and a value below 1 indicates suppression. Many practical junction devices sit close to 1, while diffusive mesoscopic conductors show a well-known universal value of 1/3. Note that the symbol F is used both for the Fano factor here and for noise factor later in this article; the two are unrelated quantities.

Spectral Characteristics

The power spectral density of flicker noise follows:

S(f) = K / f^α

Where K is a constant that depends on the device and its operating conditions, and α lies close to 1, which gives the noise its name. Measured values of α range from roughly 0.8 to 1.4 depending on the device and process, so no single exponent should be assumed. A strict 1/f law would also imply infinite power as frequency approaches zero; in practice the spectrum is only ever measured over a finite band, and the integrated noise grows merely as the logarithm of the frequency ratio, which keeps the total finite for any real observation time.

Because the density rises as frequency falls, flicker noise dominates at low frequencies. At higher frequencies the white thermal or shot contribution eventually exceeds it. The crossover, called the corner frequency or knee frequency, marks where flicker noise ceases to matter.

Amplifier datasheets normally specify this corner. A low-noise bipolar operational amplifier may place its corner near 10 Hz, a CMOS-input part typically places it higher, often above 100 Hz, and a chopper-stabilized amplifier removes the corner from the signal band altogether. Comparing corner frequencies is only meaningful alongside the white noise floor, because a device with a very low floor will show a higher corner for the same amount of flicker noise.

Physical Origins

Flicker noise arises from multiple mechanisms, which explains its ubiquity:

• Carrier trapping and release: Defects in semiconductor materials randomly capture and release charge carriers, modulating current flow
• Surface effects: Interface states between silicon and oxide trap carriers with a wide range of time constants
• Mobility fluctuations: Random variations in carrier mobility caused by lattice scattering and defects
• Contact noise: Fluctuating contact resistance at material interfaces and at resistor end caps

The 1/f spectrum emerges when many processes with different time constants contribute. Each process on its own produces a Lorentzian spectrum, flat at low frequencies and rolling off above its characteristic frequency. Superposing many such processes whose time constants are distributed logarithmically produces the characteristic 1/f shape. The two long-standing models, McWhorter's carrier-number fluctuation picture and Hooge's mobility-fluctuation picture, each explain part of the observed behavior, and modern device models draw on both.

Flicker Noise in Different Technologies

Different device technologies exhibit markedly different flicker noise:

• MOSFETs: Relatively high flicker noise, because conduction occurs at the oxide interface where traps are plentiful; noise power scales inversely with gate area, so small devices are noisier
• BJTs: Lower flicker noise than MOSFETs, with base current the primary source
• JFETs: Very low flicker noise, because current flows through bulk material away from surfaces
• Resistors: Carbon composition and thick-film resistors show high excess noise; metal film, thin film, and wirewound types are far quieter

For low-frequency precision work, JFET and bipolar input stages are therefore often preferred over plain MOSFET inputs, unless chopping or auto-zeroing removes the flicker contribution by other means.

Characterization Parameters

Flicker noise is characterized by several parameters:

• Corner frequency (f_c): The frequency at which the flicker density equals the white noise density
• Noise voltage at 1 Hz: Often quoted for operational amplifiers as the density at 1 Hz
• 0.1 Hz to 10 Hz noise: Peak-to-peak noise in this band, the standard figure of merit for precision direct-coupled applications
• Noise index: For resistors, the excess noise in decibels relative to a reference of 1 microvolt per volt of applied direct voltage, per decade of frequency

Resistors carrying direct current add excess noise on top of their thermal noise. The noise index NI expresses that excess as the RMS noise voltage generated per volt of applied direct voltage in each decade of frequency, referred in decibels to 1 microvolt per volt:

NI = 20 log₁₀(V_n / V_DC)

Here V_n is the excess noise voltage in microvolts RMS per decade and V_DC is the applied direct voltage in volts. A resistor with NI = 0 dB therefore produces 1 µV RMS of excess noise per volt of bias in each frequency decade. Carbon composition resistors are among the noisiest, with indices in the region of −10 dB to 0 dB; thick-film types typically fall between −20 dB and −10 dB; metal film types commonly run from −40 dB to −20 dB; and thin-film and bulk metal foil types fall below −40 dB, where excess noise is negligible against thermal noise. Excess noise vanishes when the resistor carries no direct current, which is why it is absent from purely alternating-signal paths.

Mitigating Flicker Noise

Several techniques reduce the impact of flicker noise:

• Larger device geometries: Flicker noise power scales inversely with device area, so larger input transistors are quieter at low frequencies
• Chopper stabilization: Modulates the signal up to a frequency where white noise dominates, amplifies it there, then demodulates
• Correlated double sampling: Samples a reference level and the signal in quick succession and subtracts, cancelling noise that is slow compared with the sampling interval
• Low-noise component selection: Prefer JFET or bipolar inputs over plain MOSFET inputs, and metal film over carbon composition resistors
• Reduce direct bias across resistors: Excess noise in a resistor scales with the applied direct voltage

Chopping and auto-zeroing carry their own costs. Both fold wideband noise into the signal band, raising the white noise floor, and both introduce switching residues at the chopping frequency that must be filtered.

Characteristics

Burst noise exhibits several distinctive features:

• Discrete levels: The noise switches between two, or occasionally more, distinct amplitude states
• Random timing: Transitions occur at random intervals with a characteristic average dwell time in each state
• Square-wave appearance: Individual bursts have abrupt rise and fall times
• Low-frequency content: Significant power at frequencies below the switching rate

The power spectral density of a single two-state switching process is Lorentzian:

S(f) = K / (1 + (f / f_c)²)

Where f_c is set by the average switching rate. The spectrum is flat well below f_c and falls as 1/f² well above it. In small devices the same two-state switching is observed directly as random telegraph signal noise, and a large collection of such processes with logarithmically distributed time constants is precisely what produces the 1/f spectrum described earlier.

Physical Mechanism

Burst noise arises from individual defects in semiconductor material that switch randomly between two states:

• Single trapping centers: One defect alternately captures and releases a carrier, modulating local current flow
• Metallic precipitates: Contamination in the semiconductor creates active trapping sites
• Crystal defects: Dislocations and other imperfections cause state switching
• Interface states: Defects at material boundaries produce the same effect

Because burst noise originates in individual defects rather than in bulk statistics, it is not fundamental in the way thermal noise is. Its presence points to a material or process problem, and it varies unpredictably from one device to the next within a lot. Well-controlled processes produce devices with little or no burst noise.

Impact on Circuits

Burst noise is particularly troublesome in certain applications:

• Low-frequency measurements: Its low-frequency content interferes with direct and slowly varying signals
• Audio applications: The characteristic popping is directly audible and objectionable
• Precision references: Step changes in a reference output degrade accuracy in a way that averaging does not cure
• Comparators: Near threshold, a burst can cause false triggering

Unlike thermal noise, which has a Gaussian amplitude distribution, burst noise is strongly non-Gaussian. An RMS figure describes it poorly, and its effect does not average away as the square root of observation time.

Screening and Selection

Because burst noise indicates device defects, screening can often remove it:

• Incoming inspection: Test devices for burst noise before assembly, observing each part for long enough to catch infrequent bursts
• Burn-in: Elevated-temperature operation stabilizes or eliminates some burst noise sources
• Selection by lot: When burst noise appears in a sample, treat the whole lot as suspect
• Process control: Improved manufacturing cleanliness reduces the incidence at its source

Modern semiconductor processes produce little burst noise in most devices, but it remains a concern for critical low-noise applications, and screening remains the practical defense because no datasheet parameter predicts it.

Avalanche Multiplication Process

In avalanche breakdown, a carrier accelerated by a strong electric field gains enough energy to free bound electrons from the lattice. Each freed electron creates an electron-hole pair, and both carriers can then ionize in turn. The multiplication amplifies the current:

I_out = M I_primary

Where M is the avalanche multiplication factor, which ranges from slightly above 1 to very large values depending on bias voltage and device structure.

The statistical nature of ionization adds randomness beyond simple shot noise. Uncertainty in where ionization events occur and in how many descendants each primary carrier produces both contribute, so the output noise exceeds M times the input noise.

Excess Noise Factor

Avalanche noise exceeds the shot noise of the multiplied current by an excess noise factor F, a quantity unrelated to the noise factor discussed later:

i_n = √(2qI_primaryM²FB)

Here I_primary is the unmultiplied current entering the multiplication region. McIntyre's analysis of 1966 gives the excess noise factor in terms of M and k, the ratio of the ionization coefficient of the non-injected carrier to that of the injected carrier:

F = kM + (1 − k)(2 − 1/M)

When k = 0, so that only one carrier type ionizes, F approaches 2 for large M. When k = 1, so that both types ionize equally, F approaches M and the noise grows in step with the gain. Silicon has a strongly asymmetric ionization ratio at visible and near-infrared wavelengths, with k of order 0.02 or lower, which is why silicon avalanche photodiodes achieve low excess noise. Compound semiconductors used at telecommunications wavelengths, such as InGaAs with InP or InAlAs multiplication layers, have larger k and correspondingly higher excess noise, which limits the useful gain.

The practical consequence is an optimum gain. Signal grows as M while the multiplied shot noise grows as M√F, so increasing M improves sensitivity only until the multiplied noise overtakes the fixed noise of the following amplifier.

Applications and Implications

Avalanche noise affects several device types:

• Avalanche photodiodes: Controlled multiplication amplifies weak optical signals before the preamplifier adds its own noise; excess noise bounds the achievable improvement
• Zener and avalanche reference diodes: Devices breaking down above roughly 6 V do so by avalanche rather than true Zener tunneling, and their noise limits the stability of voltage references
• Transistors: Operation near breakdown raises noise sharply, so low-noise designs keep well clear of it
• Random number generators: Avalanche noise serves as a physical entropy source for cryptographic applications

Minimizing Avalanche Noise

Where avalanche noise is unwanted:

• Operate below breakdown: Keep voltage margins sufficient to avoid avalanche conditions
• Use buried zeners: Subsurface breakdown in precision references is quieter and more stable than surface breakdown
• Select materials with low k: For avalanche photodiodes, a small ionization ratio minimizes the excess noise factor
• Limit multiplication: A lower M reduces total noise, trading against signal amplification
• Filter the output: Reference diodes benefit from adequate bypassing and filtering, at the cost of settling time

Definitions

Noise factor F is the ratio of the signal-to-noise ratio at the input to the signal-to-noise ratio at the output, with the source at the standard reference temperature T₀ = 290 K:

F = SNR_in / SNR_out

Since any real device adds noise of its own, the output signal-to-noise ratio is always lower than the input value, making F greater than or equal to 1.

Noise figure NF expresses the same quantity in decibels:

NF = 10 log₁₀F

The distinction matters in every calculation that follows. F is a plain power ratio and NF is its decibel form; substituting a decibel value into a formula that calls for the ratio is one of the most common errors in noise work. An ideal noiseless device has F = 1 and NF = 0 dB. Practical low-noise amplifiers achieve noise figures between about 0.5 dB and 3 dB, while general-purpose devices may sit at 6 dB to 10 dB or higher.

Alternative Definitions

Noise factor can equivalently be written as the ratio of total available output noise power to the part of that noise contributed by the source alone, again with the source at 290 K. The denominator represents the amplified source noise, so this form shows that noise factor compares the total output noise with the minimum the device could produce if it added nothing of its own.

For a device with available power gain G and equivalent input noise power N_a referred to its input:

F = 1 + N_a / (kT₀B)

Here kT₀B is the available thermal noise power from the source at the reference temperature. At 290 K the available noise power density kT₀ is about −174 dBm per hertz, the figure from which receiver sensitivity budgets normally start. This formulation ties noise factor directly to the device's equivalent input noise.

The reference temperature is a convention, not a measurement condition. A device does not become quieter when the laboratory is cold; the definition simply fixes the source noise against which its added noise is judged.

Cascaded Noise Figure

When stages are cascaded, the overall noise factor follows the Friis formula:

F_total = F₁ + (F₂ − 1)/G₁ + (F₃ − 1)/(G₁G₂) + ...

Every F and every G in this expression is a plain power ratio. Noise figures in decibels and gains in decibels must be converted to ratios before the formula is applied, and the result converted back with NF = 10 log₁₀F. The gains are available power gains, which means the formula assumes each stage is driven and loaded as it was when its noise factor was specified.

The equation reveals the governing design principles:

• First stage dominates: The first stage's noise factor enters the total undivided
• Later stages are divided by preceding gain: High first-stage gain suppresses the contribution of everything downstream
• Low-noise front end is essential: The quietest device with adequate gain belongs first in the chain, ahead of any lossy filter or cable run

Consider a first stage with F₁ = 2 (3.0 dB) and G₁ = 100 (20 dB), followed by a stage with F₂ = 10 (10 dB). The second stage contributes (10 − 1)/100 = 0.09, giving F_total = 2.09 and NF_total = 3.2 dB. A noisy second stage costs only 0.2 dB. Reverse the order and the total becomes far worse, which is why front-end placement dominates receiver design.

Measurement Considerations

Noise figure measurement demands attention to several factors:

• Source impedance: Noise figure depends on the impedance presented to the input, so any specification must state the source conditions
• Measurement bandwidth: The bandwidth must capture all relevant noise while excluding interference and image responses
• Temperature: Source noise and device noise both depend on temperature; the reference is 290 K
• Calibration: The measurement requires a characterized noise source and correction for the noise of the measuring receiver itself

The Y-factor method is the standard technique. Output noise power is measured with the input driven by a calibrated noise source switched between its hot and cold states, and the ratio of the two readings yields the noise figure without requiring a separate gain measurement. The noise source is characterized by its excess noise ratio, and the second-stage correction removes the analyzer's own contribution using the Friis formula.

Optimum Source Impedance

For most amplifiers, noise figure depends on the source impedance. An optimum source impedance Z_opt minimizes it:

• Below Z_opt: The device's voltage noise dominates, because the source noise falls faster than the amplifier's
• Above Z_opt: The device's current noise dominates, because it develops a larger voltage across the higher impedance
• At Z_opt: The two contributions are balanced

For an amplifier characterized by uncorrelated voltage and current noise densities, the optimum source resistance is simply R_opt = e_n / i_n. Minimum noise figure rarely coincides with maximum power transfer, so low-noise radio-frequency design usually accepts a deliberate input mismatch in exchange for lower noise.

A complete noise characterization therefore comprises four parameters: the minimum noise figure, the real and imaginary parts of the optimum source impedance, and the equivalent noise resistance, which describes how rapidly noise figure degrades as the source departs from optimum.

Definition and Calculation

For a filter with power transfer function |H(f)|² and peak response at frequency f₀:

ENBW = (1 / |H(f₀)|²) ∫₀^∞ |H(f)|² df

The integral accounts for every frequency at which the filter passes noise, including the skirts beyond the nominal cutoff. For white noise with power spectral density S₀, the total output noise power is then:

P_noise = S₀ ENBW

Equivalently, a voltage noise density e_n produces a total of e_n√(ENBW) volts RMS. This multiplication replaces an integration in ordinary design work. Note that ENBW always exceeds the −3 dB bandwidth, so using the −3 dB figure understates the noise.

ENBW for Common Filters

Standard filter types have known ENBW values relative to their −3 dB bandwidth, f_{3 dB}:

• First-order low-pass: ENBW = (π/2) f_{3 dB} ≈ 1.57 f_{3 dB}, exactly, for any single-pole response
• Second-order Butterworth: ENBW ≈ 1.11 f_{3 dB}
• Second-order Bessel: ENBW ≈ 1.15 f_{3 dB}
• Third-order Butterworth: ENBW ≈ 1.05 f_{3 dB}
• Fourth-order Butterworth: ENBW ≈ 1.03 f_{3 dB}
• Ideal brick-wall filter: ENBW = f_{3 dB}, the theoretical limit

For an n-pole Butterworth response the ratio is (π/2n) / sin(π/2n), which yields each of the values above and approaches 1 as the order rises. The single-pole factor of 1.57 is the one that matters most in practice, because a simple RC network terminates so many signal paths, and ignoring it understates the noise by more than 25 percent. Higher-order filters come closer to the brick-wall ideal but bring group delay variation, component sensitivity, and settling behavior that may not suit the application.

Practical Applications

ENBW simplifies several common calculations:

• Receiver sensitivity: Compute the minimum detectable signal from noise density and ENBW
• SNR prediction: Determine output SNR from the input signal and the noise spectral density
• Filter selection: Choose a bandwidth no wider than the signal requires, since total noise grows as its square root
• Measurement accuracy: Predict measurement uncertainty from sensor noise and instrument bandwidth

The same idea governs the noise bandwidth of a discrete Fourier transform bin, where the window function rather than an analog filter sets the shape. A rectangular window gives an ENBW of exactly one bin, while a Hann window gives 1.5 bins, and noise density estimates must be scaled accordingly.

Relationship to Resolution Bandwidth

In spectrum analyzers, the resolution bandwidth setting determines the filter through which the spectrum is swept, and the displayed noise level depends on it. For a voltage noise density e_n, the RMS noise voltage the instrument displays is:

V_noise = e_n √(ENBW)

When comparing noise measurements taken at different resolution bandwidth settings, normalize to a common bandwidth:

V₂ = V₁ √(ENBW₂ / ENBW₁)

Reducing the resolution bandwidth by a factor of 10 therefore lowers the displayed noise floor by 10 dB in power terms, while leaving a coherent signal unchanged. That asymmetry is what lets a narrow resolution bandwidth pull a carrier out of the noise. Analyzers that report a normalized noise floor in dBm per hertz have already applied this correction, together with a small allowance for the logarithmic detector's response to Gaussian noise.

Autocorrelation Function

The autocorrelation function R(τ) measures the similarity between a signal and a time-shifted copy of itself:

R(τ) = E[x(t) x(t + τ)]

Where E[ ] denotes the expected value, which for a stationary ergodic process equals a time average. Key properties include:

• R(0) equals the mean-square value: The autocorrelation at zero delay equals the noise power
• Symmetric about zero: R(−τ) = R(τ) for real signals
• Maximum at zero: R(0) ≥ |R(τ)| for all τ
• Approaches zero: For most noise, R(τ) decays as the delay grows

The correlation time, the delay at which R(τ) has fallen substantially below R(0), characterizes how quickly noise fluctuations become independent. Wide bandwidth implies a short correlation time, which is why white noise is uncorrelated from sample to sample.

Power Spectral Density

The power spectral density S(f) describes noise power per unit frequency. It is related to the autocorrelation function by the Wiener-Khinchin theorem, which states that one is the Fourier transform of the other:

S(f) = ∫ R(τ) e^−j2πfτ dτ

Conversely, R(τ) is the inverse Fourier transform of S(f). This relationship connects the time-domain and frequency-domain descriptions of noise, and it is what allows a measurement made in either domain to predict behavior in the other.

Common spectral shapes include:

• White noise: S(f) is constant, and R(τ) is an impulse at τ = 0
• Flicker noise: S(f) ∝ 1/f, and R(τ) carries a logarithmic dependence with no finite correlation time
• Lorentzian: S(f) = K / (1 + (f/f_c)²), and R(τ) = exp(−|τ| / τ_c), where τ_c = 1/(2πf_c)

Engineering practice uses the single-sided density, defined for positive frequencies only, which is why the factors of 4 and 2 appear in the thermal and shot noise formulas. The two-sided density used in theoretical work is half as large at each frequency.

Cross-Correlation

Cross-correlation measures the relationship between two different signals:

R_xy(τ) = E[x(t) y(t + τ)]

For independent noise sources the cross-correlation is zero at every delay. A non-zero cross-correlation indicates one of several things:

• Common noise sources: Both signals contain components from the same physical source
• Correlated processes: The noise-generating mechanisms are related
• Signal components: A deterministic signal is present in both channels

Cross-correlation analysis helps identify noise sources, separate signal from noise, and recover signals buried below the noise floor. The cross-correlation measurement technique exploits this directly: two independent amplifiers observe the same source, and averaging their cross-spectrum suppresses each amplifier's own noise while preserving the common source noise, allowing measurements below the noise floor of either instrument alone.

Cross Power Spectral Density

The cross power spectral density S_xy(f) is the Fourier transform of the cross-correlation function. It reveals how two signals share frequency content:

• Magnitude: Indicates the shared power at each frequency
• Phase: Shows the phase relationship between the shared components
• Coherence: The normalized magnitude squared, between 0 and 1, indicates how linearly related the two signals are at each frequency

Coherence analysis distinguishes correlated noise, which arrives from a common source and may be cancelled, from uncorrelated noise, which may only be reduced by averaging. A coherence near 1 across a band points to a shared origin worth hunting down, such as a common supply rail or ground path.

Practical Spectrum Analysis

Measuring noise spectra requires appropriate technique:

• Averaging: Many spectral estimates are averaged to reduce the scatter in the estimate
• Windowing: An appropriate window function reduces spectral leakage between bins
• Resolution-bandwidth trade-off: Finer frequency resolution requires a longer observation time
• Anti-aliasing: Filtering before digitization prevents out-of-band noise from folding into the measurement band

A single periodogram is a poor estimator: its standard deviation equals the quantity being estimated, regardless of record length. Averaging N independent estimates reduces the variance of the result by a factor of N, so the standard deviation falls as 1/√N. Reducing the scatter tenfold therefore requires a hundred averages, which is why low-frequency noise measurements take so long.

Definition of Noise Temperature

The noise temperature T_n of a source is defined so that a matched resistor at temperature T_n would deliver the same available noise power:

P_n = kT_nB

The definition assumes a white spectrum over the bandwidth of interest. For a source delivering available noise power P_n in bandwidth B:

T_n = P_n / (kB)

Noise temperature is a bookkeeping device rather than a physical temperature. It may far exceed any physical temperature in the system, as it does for a calibrated noise diode, and it may fall well below the ambient temperature for a cooled low-noise amplifier.

Relationship to Noise Figure

Noise temperature and noise factor are two statements of the same information:

T_e = T₀(F − 1)

Where T_e is the equivalent noise temperature of the device and T₀ is the reference temperature, 290 K by IEEE convention. Inverting:

F = 1 + T_e / T₀

At T₀ = 290 K:

• F = 2 (NF = 3.01 dB) corresponds to T_e = 290 K
• F = 1.5 (NF = 1.76 dB) corresponds to T_e = 145 K
• F = 1.1 (NF = 0.41 dB) corresponds to T_e = 29 K

Noise temperature is preferred for very low-noise systems because it resolves differences that the decibel scale compresses. The step from 0.41 dB to 0.28 dB looks trivial, yet it represents a fall from 29 K to 19 K, a third of the receiver's noise contribution.

System Noise Temperature

For a complete receiving system, the system noise temperature includes contributions from the antenna as well as the receiver:

T_sys = T_ant + T_rx

Where T_ant is the antenna noise temperature, determined by what the antenna sees, and T_rx is the receiver noise temperature referred to the same reference plane.

The antenna temperature gathers several contributions:

• Sky noise: Thermal emission from the atmosphere, plus the cosmic microwave background at about 2.7 K
• Ground noise: Emission from the earth picked up through sidelobes, near the physical ground temperature
• Ohmic losses: Resistive loss in the antenna structure, which both attenuates the signal and adds thermal noise
• Feed line losses: Cable or waveguide loss ahead of the first amplifier

Receiver sensitivity depends on the total system noise temperature, not on the receiver alone. Once T_ant dominates, further improvement of the receiver buys nothing, which is why terrestrial links and satellite earth stations set very different targets for receiver noise temperature.

Cascaded Noise Temperature

For cascaded stages, noise temperatures combine as:

T_total = T₁ + T₂/G₁ + T₃/(G₁G₂) + ...

This is the Friis formula expressed in temperature rather than noise factor, and as before the gains are power ratios, not decibels. The first stage's noise temperature enters undivided, and later stages are divided by all the gain ahead of them.

Consider a cryogenic low-noise amplifier with T₁ = 10 K and 30 dB of gain, which is a power ratio of 1000, followed by a room-temperature stage with T₂ = 290 K:

T_total = 10 K + 290 K / 1000 = 10.29 K

The warm second stage contributes about 3 percent of the total. Cutting the first-stage gain to 20 dB would raise the total to 12.9 K, so front-end gain is as important as front-end noise.

Noise Temperature of Attenuators

A passive attenuator, or any lossy component, at physical temperature T has an equivalent noise temperature:

T_atten = T(L − 1)

Where L is the loss factor, the ratio of input power to output power, so that L ≥ 1. A matched attenuator at 290 K has a noise figure numerically equal to its loss in decibels, a result that follows directly from thermodynamic equilibrium.

Placed ahead of an amplifier, an attenuator degrades system noise temperature twice over:

• It adds its own noise contribution
• It reduces the signal level, so the following stage weighs more heavily
• Combining both effects: T_sys = T(L − 1) + L T_amp

A 3 dB cable loss ahead of a 50 K amplifier at 290 K ambient therefore yields 290(2 − 1) + 2 × 50 = 390 K, nearly eight times the amplifier's own figure. This is why a low-noise amplifier belongs at the antenna rather than at the far end of the feed line.

Cryogenic Cooling

For the most demanding applications, amplifier noise temperature can be reduced by cooling:

• Liquid nitrogen (77 K): Practical for many applications, and it reduces the thermal contribution of resistive elements by roughly a factor of four relative to 290 K
• Liquid helium (4.2 K): A further large reduction at considerably greater complexity; long used in radio astronomy
• Closed-cycle coolers: No consumable cryogen, with physical temperatures near 15 K to 20 K routinely achieved
• Dilution refrigerators: Millikelvin temperatures for quantum-limited devices and superconducting detectors

Cooling reduces the thermal noise of resistive elements and improves the noise performance of active devices, particularly high electron mobility transistors, whose channel noise falls with physical temperature. The improvement is not simply proportional to temperature, because some noise mechanisms are only weakly temperature dependent, and every cooled front end must still contend with the loss of the window, feed, and connections ahead of it. At the extreme, quantum mechanics imposes a floor: a phase-insensitive linear amplifier cannot have a noise temperature below about hf/k, which is roughly 0.5 K at 10 GHz.