Noise and Basic Filtering

Experimental noise models, signal-to-noise ratio, moving averages, and introductory low-pass filtering.

What it is

Physical noise is any undesired stochastic fluctuation superimposed on an experimental signal that obscures information or degrades precision. Far from being a manufacturing defect, noise is a direct physical manifestation of the atomic, discrete, and thermodynamic nature of matter.

Fundamental noise mechanisms are classified by their physical origin:

1. Johnson-Nyquist thermal noise: Generated by random thermal Brownian motion of conduction electrons inside any dissipative resistor $R$. It is a white noise (flat power spectral density) governed by the fluctuation-dissipation theorem: $S_v(f) = 4 k_B T R \quad [\text{V}^2/\text{Hz}], \qquad V_{\text{rms}} = \sqrt{4 k_B T R \Delta f}$
2. Shot noise (Schottky noise): Driven by the quantized nature of electrical charge ($e$) and the Poisson arrival statistics of carriers traversing potential barriers (p-n junctions, photodetectors): $S_i(f) = 2 e I \quad [\text{A}^2/\text{Hz}], \qquad I_{\text{rms}} = \sqrt{2 e I \Delta f}$
3. Flicker noise ($1/f$ noise): Predominates at low frequencies, originating from stochastic charge carrier trapping in semiconductor surface states or slow ambient thermal drift.
4. Physical filtering: Passing the signal through a Linear Time-Invariant (LTI) system—analog or digital—whose frequency response $H(f)$ selectively attenuates spectral bands dominated by noise while passing signal harmonics.

Why it appears in physics

No experimental measurement in physics or astrophysics operates in a zero-noise regime, from gravitational wave interferometry in LIGO to nuclear magnetic resonance and scanning tunneling microscopy:

• Noise establishes the fundamental thermodynamic detection floor (standard quantum limit or thermal noise limit).
• Without adequate filtering, wideband noise obscures signals; during digital analog-to-digital conversion (ADC), out-of-band noise folds back into the baseband (aliasing) whenever Nyquist's criterion ($f_s \geq 2 f_{\text{max}}$) is violated.
• Analog RC filters, moving-average filters, and digital Butterworth/Chebyshev filters enhance the signal-to-noise ratio (SNR) by orders of magnitude, turning buried traces into statistically robust physical discoveries.

Visual or geometric idea

In the time domain, a noise-corrupted trace resembles a ragged trajectory with erratic spikes masking the smooth physical path. In the frequency domain (Power Spectral Density, PSD):

• White noise appears as a flat horizontal baseline extending across all frequencies.
• $1/f$ noise rises steeply towards the origin (DC limit).
• Physical signals concentrate in discrete spectral peaks or compact bandwidths. Geometrically, a low-pass filter functions as a spectral gate cutting off frequencies above cutoff $f_c$, pruning the vast high-frequency noise floor while preserving the energy inside the signal envelope.

Minimal symbolic language

1. Johnson-Nyquist thermal noise voltage: $V_n = \sqrt{4 k_B T R \Delta f}$
2. Schottky shot noise current: $I_n = \sqrt{2 e I \Delta f}$
3. Signal-to-Noise Ratio (SNR) in decibels: $\text{SNR} = 10 \log_{10}\left( \frac{P_{\text{signal}}}{P_{\text{noise}}} \right) = 20 \log_{10}\left( \frac{V_{\text{signal, rms}}}{V_{\text{noise, rms}}} \right)$
4. Passive RC low-pass filter cutoff frequency: $f_c = \frac{1}{2\pi R C}, \qquad |H(f)| = \frac{1}{\sqrt{1 + (f/f_c)^2}}$
5. $M$-point moving-average digital filter: $y[n] = \frac{1}{M} \sum_{k=0}^{M-1} x[n-k], \qquad \sigma_y = \frac{\sigma_x}{\sqrt{M}}$

What the result means

The computed $V_{\text{rms}}$ defines the standard deviation of random voltage variations added to any single measurement. Applying an ideal band-limiting filter with noise equivalent bandwidth $\Delta f$:

• Reduces white noise variance in direct proportion to bandwidth narrowing: $\sigma_{\text{filtered}}^2 = S_0 \Delta f$.
• For an $M$-point moving-average filter across uncorrelated samples, noise standard deviation drops by exactly $1/\sqrt{M}$. Averaging 100 successive snapshots cuts noise by a factor of 10 ($90\%$ reduction).

When not to use it

1. Filtering without considering response time or phase distortion: Setting low-pass cutoff $f_c$ too low attenuates high-frequency signal components, broadening rapid physical transients and introducing group delays.
2. Moving-average filters on fast oscillating waveforms: For narrowband signals or fast transient pulses, simple moving averages act as sinc filters that severely distort wave profiles.
3. Omitting analog anti-aliasing prior to sampling: Digital filtering cannot rescue signals sampled below the Nyquist rate; high-frequency noise folds permanently into baseband during digitization. Anti-aliasing filtering must be performed in the analog domain before ADC conversion.

Operational summary

• Identify the dominant physical noise source: thermal (resistors), shot noise (junction currents), or environmental (50/60 Hz mains interference).
• Calculate the system's noise equivalent bandwidth $\Delta f$.
• Map the spectral frequency band containing the target physical signal.
• Design a filter whose cutoff $f_c$ separates signal from out-of-band noise without clipping critical harmonics.
• Employ $M$-sample averaging to reduce random noise by $\sqrt{M}$ whenever signals are quasi-stationary.

Physical Noise Characterization and Filter Design Procedure

Before calculating

Before selecting signal conditioning components or digital filtering routines:

• Measure or extract physical sensor parameters: absolute temperature $T$ (kelvins), equivalent Thevenin source resistance $R$ (ohms), and DC bias current $I_{\text{DC}}$ (amperes).
• Define maximum physical signal frequency $f_{\text{signal, max}}$ to establish required bandwidth without losing dynamic content.

Step 1: identify the mathematical object

Formulate power spectral densities for dominant physical noise sources:

1. Thermal noise: $S_v = 4 k_B T R \quad [\text{V}^2/\text{Hz}], \qquad k_B = 1.380649 \times 10^{-23} \text{ J/K}$
2. Shot noise: $S_i = 2 e I_{\text{DC}} \quad [\text{A}^2/\text{Hz}], \qquad e = 1.60217663 \times 10^{-19} \text{ C}$
3. Combined input-referred noise voltage: $V_{n,\text{total}}^2 = S_v + S_i \cdot R^2 = (4 k_B T R + 2 e I_{\text{DC}} R^2) \Delta f$

Step 2: choose the operation

Select the appropriate filtering strategy:

• Broadband white noise reduction: Low-pass filter with cutoff $f_c \approx 1.5 \times f_{\text{signal, max}}$.
• Mains line-frequency interference ($50\text{ Hz} / 60\text{ Hz}$): Twin-T or digital notch filter.
• Sampled stationary digital traces: Moving-average or Savitzky-Golay polynomial filter.

Step 3: preserve units and signs

• Boltzmann constant $k_B$ in SI units is $\text{J/K} = \text{V}^2/(\Omega\cdot\text{Hz}\cdot\text{K})$.
• $4 k_B T R$ carries exact dimensions of $\text{V}^2/\text{Hz}$.
• Taking the square root $\sqrt{S_v}$ yields amplitude spectral density in $\text{V}/\sqrt{\text{Hz}}$ or $\text{nV}/\sqrt{\text{Hz}}$.
• Multiply by the square root of noise equivalent bandwidth $\sqrt{\Delta f}$ (in $\sqrt{\text{Hz}}$) to obtain root-mean-square noise in volts ($V_{\text{rms}}$).

Step 4: calculate

1. RMS noise voltage over filtered bandwidth: $V_{\text{rms}} = \sqrt{(4 k_B T R + 2 e I_{\text{DC}} R^2) \cdot \frac{\pi}{2} f_c}$ (Note: for a first-order passive RC filter, equivalent noise bandwidth is $\Delta f_{\text{ENB}} = \frac{\pi}{2} f_c \approx 1.571 f_c$ due to the gradual 20 dB/decade roll-off).
2. Signal-to-Noise Ratio: $\text{SNR}_{\text{dB}} = 20 \log_{10}\left( \frac{V_{\text{signal, rms}}}{V_{\text{rms}}} \right)$
3. SNR gain through $M$-sample averaging: $\Delta \text{SNR}_{\text{dB}} = 10 \log_{10}(M)$

Step 5: interpret the result

• Contrast pre- and post-filtering SNR: a $20\text{ dB}$ improvement represents a factor-of-10 reduction in oscilloscope peak-to-peak noise fuzz.
• Check filter rise time $t_r \approx \frac{0.35}{f_c}$: filter response time must remain noticeably faster than physical transient steps to prevent distortion.

Self-check checklist

• Did I convert temperature to kelvins ($T = T_C + 273.15$)?
• Do noise spectral densities carry units of $\text{V}/\sqrt{\text{Hz}}$ before bandwidth integration?
• Did I account for the $\pi/2$ equivalent noise bandwidth factor for single-pole RC filters?
• Is the ADC sampling frequency at least double the highest spectral component to prevent aliasing?
• Did I confirm that filter smoothing does not round off genuine physical transients?

Guided Example: Suppressing Thermal Noise in a Piezoelectric Pressure Transducer

Physical Situation

A high-impedance piezoelectric pressure transducer ($R = 10.0 \,\text{k}\Omega$) operates at ambient room temperature ($T = 293.15 \text{ K}$) to measure combustion acoustic oscillations with physical frequency content bounded below $f_{\max} = 2.0 \text{ kHz}$. Expected signal amplitude is $V_{\text{signal}} = 1.0 \text{ mV}$ RMS ($1000 \,\mu\text{V}$). Unfiltered data acquisition spans a broad bandwidth $\Delta f_0 = 10.0 \text{ MHz}$ ($10^7 \text{ Hz}$). The goal is to:

1. Compute the broadband Johnson-Nyquist noise voltage $V_{n,0}$ and initial SNR.
2. Design a passive first-order RC low-pass filter suited to the acoustic signal.
3. Compute the filtered noise voltage $V_{n,1}$ and net SNR improvement.

Mathematical Translation

1. Initial Thermal Noise: $V_{n,0} = \sqrt{4 k_B T R \Delta f_0}$ where $k_B = 1.38065 \times 10^{-23} \text{ J/K}$, $T = 293.15 \text{ K}$, $R = 10^4 \,\Omega$.
2. RC Filter Design: To avoid attenuating signal content below $2.0 \text{ kHz}$, set cutoff to $f_c = 2.5 \text{ kHz}$: $f_c = \frac{1}{2\pi R C} \implies C = \frac{1}{2\pi R f_c}$
3. Equivalent Noise Bandwidth (ENB): $\Delta f_{\text{ENB}} = \frac{\pi}{2} f_c = \frac{\pi}{2} (2500) \approx 3927 \text{ Hz}$
4. Filtered Thermal Noise: $V_{n,1} = \sqrt{4 k_B T R \Delta f_{\text{ENB}}}$

Step-by-Step Calculation

1. Noise Spectral Density: $S_v = 4 k_B T R = 4 \times (1.38065 \times 10^{-23}) \times 293.15 \times 10^4 = 1.6186 \times 10^{-16} \text{ V}^2/\text{Hz}$ $\sqrt{S_v} \approx 1.272 \times 10^{-8} \text{ V}/\sqrt{\text{Hz}} = 12.72 \text{ nV}/\sqrt{\text{Hz}}$
2. Unfiltered Broadband Noise ($\Delta f = 10^7 \text{ Hz}$): $V_{n,0} = 1.272 \times 10^{-8} \times \sqrt{10^7} \approx 40.23 \,\mu\text{V}$ $\text{SNR}_0 = 20 \log_{10}\left( \frac{1000 \,\mu\text{V}}{40.23 \,\mu\text{V}} \right) \approx 27.91 \text{ dB}$
3. Filter Capacitor: $C = \frac{1}{2\pi \times 10000 \times 2500} \approx 6.37 \text{ nF}$ (Standard $6.8 \text{ nF}$ ceramic component).
4. Filtered Noise: $V_{n,1} = 1.272 \times 10^{-8} \times \sqrt{3927} \approx 0.80 \,\mu\text{V}$ $\text{SNR}_1 = 20 \log_{10}\left( \frac{1000 \,\mu\text{V}}{0.80 \,\mu\text{V}} \right) \approx 61.94 \text{ dB}$

Reading the Result

Passive RC filtering suppresses thermal noise from $40.2 \,\mu\text{V}$ down to $0.80 \,\mu\text{V}$ (a 50-fold reduction in RMS fluctuations), boosting signal-to-noise ratio by $\Delta \text{SNR} = 34.03 \text{ dB}$. On an oscilloscope, the blurry $4\%$ baseline trace sharpens into a clean trajectory with under $0.08\%$ residual noise.

Common Error Avoided

Directly taking cutoff frequency $f_c$ as noise bandwidth $\Delta f$ while neglecting the single-pole roll-off tail ($-20\text{ dB/decade}$), which requires multiplying by the analytical correction factor $\pi/2 \approx 1.57$.

Inverse Example: Absolute Thermometry via Johnson Noise Voltage Measurements

Physical Situation

Johnson noise thermometry provides a primary thermodynamic temperature standard independent of empirical mercury or thermocouple calibrations. Thermal noise across an ultra-stable calibrated resistor $R = 100.00 \,\text{k}\Omega$ is recorded using a cryogenic low-noise preamplifier through a flat bandpass filter between $f_1 = 10.00 \text{ kHz}$ and $f_2 = 20.00 \text{ kHz}$. An RMS voltmeter measures an output voltage fluctuation of $V_{\text{rms}} = 1.282 \,\mu\text{V}$.

Given Result

Measured RMS thermal noise: $V_{\text{rms}} = 1.282 \times 10^{-6} \text{ V}$. Calibrated resistor: $R = 1.000 \times 10^5 \,\Omega$. Rectangular bandwidth: $\Delta f = 1.000 \times 10^4 \text{ Hz}$.

Analytical Reconstruction

Nyquist's foundational theorem states: $V_{\text{rms}}^2 = 4 k_B T R \Delta f$ Solving for absolute thermodynamic temperature $T$: $T = \frac{V_{\text{rms}}^2}{4 k_B R \Delta f}$ Substituting numerical values: $V_{\text{rms}}^2 = 1.643524 \times 10^{-12} \text{ V}^2$ $4 k_B R \Delta f = 4 \times (1.380649 \times 10^{-23}) \times 10^5 \times 10^4 = 5.522596 \times 10^{-14} \text{ V}^2/\text{K}$ $T = \frac{1.643524 \times 10^{-12}}{5.522596 \times 10^{-14}} \approx 297.60 \text{ K}$ In Celsius: $T_C = 297.60 - 273.15 = 24.45^\circ\text{C}$

What Physical Question It Can Answer

Why does this technique yield absolute temperature inside nuclear reactors or helium cryostats without secondary sensors? Because Boltzmann's constant $k_B$ is an exact universal physical constant in SI units. Measuring electron thermal variance directly probes average thermal energy per thermodynamic degree of freedom $\frac{1}{2} k_B T$.

Boundary Verification

As $T \to 0 \text{ K}$, classical theory predicts $V_{\text{rms}} \to 0$. In the ultra-cold quantum regime ($k_B T \ll h f$), the Callen-Welton quantum fluctuation-dissipation theorem replaces $4 k_B T$ with $2 h f \coth(h f / 2 k_B T)$, preserving quantum zero-point fluctuations $\frac{1}{2} h f$.

Common Mistakes in Physical Noise and Signal Filtering

1. Attempting to cure aliasing via post-ADC digital filtering

It seems reasonable because software digital filters are flexible, bug-free, and easily tuned in code. It fails because once a continuous signal is digitized below the Nyquist threshold ($f_s < 2 f_{\text{noise}}$), out-of-band high frequencies alias irreversibly into the baseband. In the digital discrete domain, folded noise is mathematically indistinguishable from authentic low-frequency physics. It is detected when slow phantom oscillations appear that alter their frequency when changing the ADC clock rate. It is corrected by always placing an analog low-pass filter prior to the ADC (anti-aliasing filter). Mini-example: sampling a $990 \text{ Hz}$ noise pickup at $f_s = 1000 \text{ Hz}$ creates a phantom $10 \text{ Hz}$ sine wave; no low-pass digital filter passing $10 \text{ Hz}$ can separate it.

2. Neglecting Equivalent Noise Bandwidth (ENB) in analog filters

It seems reasonable because the $-3\text{ dB}$ cutoff frequency $f_c$ is widely taught as the filter bandwidth. It fails because analog circuits do not roll off as step functions. A 1st-order RC filter rolls off at $-20\text{ dB/decade}$, allowing white noise integrated beyond $f_c$ to contribute $57\%$ extra variance: $\Delta f_{\text{ENB}} = \frac{\pi}{2} f_c \approx 1.571 f_c$. It is detected when experimental noise measurements exceed calculations based on $\sqrt{4 k_B T R f_c}$ by roughly $25\%$. It is corrected by applying the appropriate ENB correction factor for the filter topology. Mini-example: for $f_c = 10 \text{ kHz}$, the effective bandwidth collecting thermal noise is $15.71 \text{ kHz}$, not $10 \text{ kHz}$.

3. Excessive bandwidth narrowing causing dynamic distortion

It seems reasonable because narrowing bandwidth steadily lowers total integrated thermal noise ($V_n \propto \sqrt{\Delta f}$). It fails because physical filters introduce phase shifts and finite response times. Setting $f_c$ too close to signal frequencies clips fast harmonics, blunting transient edges and introducing phase distortion. It is detected when filtered signal rise time matches the filter's own characteristic rise time $t_r \approx 0.35 / f_c$. It is corrected by maintaining $f_c$ at least $1.5$ to $2$ times higher than the highest essential harmonic of the physical phenomenon. Mini-example: filtering a $1 \text{ ms}$ collision impact transient with a $50 \text{ Hz}$ filter flattens the shock peak into a distorted, delayed bump.

4. Conflating Johnson thermal noise with Schottky shot noise

It seems reasonable because both manifest as white noise (flat power spectral density) across most common frequencies. It fails because underlying physical mechanisms and parameter dependencies are fundamentally different: thermal noise scales with resistance and absolute temperature ($4 k_B T R$) and exists under zero current; shot noise scales with DC net current ($2 e I$) and is temperature-independent. It is detected when experimentalists attempt to suppress shot noise by cryogenic cooling, which produces zero change in Schottky fluctuations. It is corrected by determining whether noise stems from dissipative resistance or carrier quantum transit across potential barriers. Mini-example: an unbiased resistor carries zero shot noise, yet exhibits measurable Johnson thermal noise.

5. Assuming moving averages reduce all forms of noise

It seems reasonable because taking averages is the common universal recipe for cleaning noisy traces. It fails because an $M$-sample average reduces noise by $1/\sqrt{M}$ only when successive samples are statistically independent (uncorrelated). When noise is dominated by $1/f$ flicker noise, thermal drift, or periodic mains hum, samples share strong autocorrelation; lengthening averaging windows hits Allan variance floors and increases error. It is detected by calculating the Allan deviation across averaging times: when the curve stops dropping as $1/\sqrt{\tau}$ and turns upward, averaging is counterproductive. It is corrected by employing AC modulation and lock-in amplification to shift signals away from the $1/f$ noise band. Mini-example: averaging for an hour across a laboratory thermistor drifts with building HVAC temperature changes, amplifying systemic measurement bias.

Quick detection rule

• Before averaging: check if noise is white. If drift or hum is present, averaging will not save you.
• Before digitizing: check for an analog front-end filter. Without it, aliasing permanently corrupts your dataset.

Physical Applications of Physical Noise and Signal Filtering

1. Gravitational wave astrophysics (LIGO and Virgo)

Appears in kilometer-scale laser interferometers measuring spacetime metric strains $h \sim 10^{-21}$ (mirror displacements below one thousandth of a proton radius, $\Delta L \sim 10^{-18} \text{ m}$). It is calculated by analyzing cross-spectral noise floors dominated by seismic ground noise at low frequencies ($< 10 \text{ Hz}$), Brownian thermal dissipation in mirror dielectric coatings at intermediate frequencies ($10 - 100 \text{ Hz}$), and photon shot noise at high frequencies ($> 100 \text{ Hz}$). The result produces the strain sensitivity curve $S_h(f)$, defining the cosmological reach for detecting binary black hole and neutron star inspirals. The critical danger is raising laser optical power to reduce photon shot noise without accounting for quantum radiation pressure fluctuations that exert mechanical recoil forces shaking test mass mirrors at low frequencies (quantum back-action). As a final sanity check, frequency-dependent squeezed vacuum states are injected to reduce phase noise without inflating radiation pressure noise.

2. Scanning Tunneling Microscopy (STM): picoampere shot noise limits

Appears in surface physics and nanoscience when scanning atomic metal tips within sub-nanometer gaps to detect quantum tunneling currents $I_t \sim 100 \text{ pA} - 1 \text{ nA}$. It is calculated by noting that for DC currents of $100 \text{ pA}$, Schottky shot noise is $I_{\text{shot}} = \sqrt{2 e I \Delta f} = \sqrt{2 \times (1.6 \times 10^{-19}) \times 10^{-10} \times 1000} \approx 0.18 \text{ pA}$ over a $1 \text{ kHz}$ bandwidth. The result sets the signal-to-noise ratio in the piezoelectric feedback loop, dictating whether individual surface adatoms can be resolved with sub-picometer vertical stability. The critical danger is allowing ambient acoustic vibrations or tip cable capacitance to couple; a mere $0.1 \text{ nm}$ mechanical shudder swings tunneling current by a factor of 10 due to exponential distance coupling $I_t \propto e^{-2\kappa z}$. As a final sanity check, the ultra-high vacuum chamber is levitated on pneumatic isolation tables and current is filtered below tip mechanical resonance frequencies ($> 5 \text{ kHz}$).

3. Nuclear Magnetic Resonance (NMR): cryogenic preamplifiers and quadrature detection

Appears in structural biology and chemistry when measuring faint free-induction decay (FID) signals emitted by proton ($^1\text{H}$) nuclear magnetic moments after radiofrequency tipping pulses. It is calculated by modeling the receiver coil as a resonant $LC$ tank. Dominant noise is Johnson-Nyquist thermal noise from coil wire resistance and dielectric specimen losses: $V_n = \sqrt{4 k_B T R \Delta f}$. The result establishes NMR spectrum SNR; cooling receiver coils with liquid helium to $T = 20 \text{ K}$ (cryoprobes) reduces thermal noise by $\sqrt{293/20} \approx 3.8$, accelerating multidimensional protein structural determinations by more than an order of magnitude in acquisition time. The critical danger is using excess receiver bandwidth or omitting in-phase/quadrature (I/Q) low-pass filtering, which folds the image noise band over the resonance line (incurring a $3\text{ dB}$ SNR penalty). As a final sanity check, matched exponential apodization windowing is applied to the time-domain FID before Fast Fourier Transformation (FFT).

4. Radio astronomy: radiometer equation and antenna temperature calibration

Appears in millimeter and centimeter radio astronomy when detecting faint cosmic sources (such as the Cosmic Microwave Background or interstellar carbon monoxide clouds at $115 \text{ GHz}$) via parabolic dishes. It is calculated using the Dicke radiometer equation: the minimum detectable antenna temperature fluctuation is $\Delta T_{\text{rms}} = \frac{T_{\text{sys}}}{\sqrt{\Delta f \cdot \tau}}$, where $T_{\text{sys}}$ is total system noise temperature, $\Delta f$ is receiver bandwidth, and $\tau$ is integration time. The result determines integration limits needed to extract faint molecular lines above receiver thermal backgrounds. The critical danger is extending integration time $\tau$ indefinitely under the assumption that noise continues scaling as $1/\sqrt{\tau}$, ignoring receiver gain drift and atmospheric water vapor fluctuations that introduce $1/f$ flicker noise plateaus for $\tau > 10 - 100 \text{ s}$. As a final sanity check, rapid beam switching or Dicke chopping alternates telescope pointing between on-source and blank sky at rates above the $1/f$ knee frequency.

5. Low-temperature physics: primary cryogenic Johnson noise thermometry

Appears in ultra-low temperature physics ($^3\text{He}-^4\text{He}$ dilution refrigerators) to measure absolute thermodynamic temperatures in the millikelvin regime ($10 \text{ mK} - 1 \text{ K}$). It is calculated from mean-squared thermal voltage fluctuations in an uncalibrated metallic sensor: $\langle V_n^2 \rangle = 4 k_B T R \Delta f$. The result supplies a primary thermodynamic scale that calibrates secondary germanium and ruthenium oxide resistance thermometers without external references. The critical danger is sensor self-heating from preamplifier bias or ambient radiofrequency pickup (RF heating), which decouples conduction electrons from the crystal phonon bath below $50 \text{ mK}$. As a final sanity check, metal powder filters and pi-filters are installed along cryogenic wiring harnesses to eliminate microwave parasitic heating above $100 \text{ MHz}$.

Guided Practice: Physical Noise and Signal Filtering

Level 1: recognition

Identify the dominant physical noise mechanism in each sensor configuration:

1. A $10 \,\text{M}\Omega$ metal film resistor connected to the open-circuit non-inverting input of an operational amplifier.
2. A photomultiplier tube or avalanche photodiode detecting a faint, stream of individual photons.
3. A bipolar junction transistor audio preamplifier operating at sub-audio frequencies ($1 - 10 \text{ Hz}$).
4. An unshielded inductive pickup placed adjacent to a $220 \text{ V}$, $50 \text{ Hz}$ power transformer.

Solution:

1. Johnson-Nyquist thermal noise (driven by high ohmic resistance).
2. Schottky shot noise (driven by discrete Poisson arrival of electrons and photons).
3. Flicker noise or $1/f$ noise (slow carrier trapping fluctuations in semiconductors).
4. Environmental electromagnetic line interference (external pickup, not an intrinsic fundamental noise floor).

Level 2: direct application

A reverse-biased photodiode generates a DC photocurrent of $I_{\text{DC}} = 5.0 \,\mu\text{A} = 5.0 \times 10^{-6} \text{ A}$. The transimpedance detection bandwidth is $\Delta f = 100 \text{ kHz} = 10^5 \text{ Hz}$. Compute RMS shot noise current $I_{\text{shot, rms}}$ and the relative noise percentage.

Solution:

1. Shot noise power spectral density: $S_i = 2 e I_{\text{DC}} = 2 \times (1.602 \times 10^{-19}) \times (5.0 \times 10^{-6}) = 1.602 \times 10^{-24} \text{ A}^2/\text{Hz}$
2. RMS shot current: $I_{\text{shot, rms}} = \sqrt{S_i \Delta f} = \sqrt{1.602 \times 10^{-24} \times 10^5} = \sqrt{1.602 \times 10^{-19}} \approx 4.002 \times 10^{-10} \text{ A} = 0.400 \text{ nA}$
3. Relative noise ratio: $\frac{I_{\text{shot, rms}}}{I_{\text{DC}}} = \frac{4.002 \times 10^{-10} \text{ A}}{5.0 \times 10^{-6} \text{ A}} \approx 8.004 \times 10^{-5} = 0.0080\% \quad (80 \text{ ppm})$

Level 3: tool selection

An electrocardiogram (ECG) trace suffers from two simultaneous noise sources:

• Broadband thermal noise above $150 \text{ Hz}$.
• A persistent $50 \text{ Hz}$ sinusoidal hum induced by laboratory mains cabling. What cascading filter topology best cleans this ECG trace (whose diagnostic bandwidth spans $0.5 \text{ Hz}$ to $100 \text{ Hz}$)?

Solution: Implement a two-stage cascade filter:

1. A low-pass filter (such as a 4th-order Butterworth) with cutoff $f_c = 100 \text{ Hz}$ to reject wideband thermal noise above the clinical diagnostic band.
2. A notch rejection filter tuned precisely to $f_0 = 50.0 \text{ Hz}$ with high quality factor ($Q \geq 10$) to notch out mains hum without eroding adjacent cardiac QRS harmonics.

Level 4: physical interpretation

You record $N$ repeated measurements of a constant physical observable corrupted by zero-mean white Gaussian noise with standard deviation $\sigma = 10.0 \text{ mV}$. Prove how many consecutive independent observations must be averaged to reduce sample mean uncertainty below $0.5 \text{ mV}$.

Solution: The sample mean estimator over $N$ independent observations is $\bar{X} = \frac{1}{N} \sum_{i=1}^N X_i$. Variance of the sample mean is: $\operatorname{Var}(\bar{X}) = \frac{1}{N^2} \sum_{i=1}^N \operatorname{Var}(X_i) = \frac{\sigma^2}{N}$ Standard error is: $\sigma_{\bar{X}} = \frac{\sigma}{\sqrt{N}}$ Enforcing $\sigma_{\bar{X}} \leq 0.5 \text{ mV}$: $\frac{10.0}{\sqrt{N}} \leq 0.5 \implies \sqrt{N} \geq 20 \implies N \geq 400$ Physical interpretation: Gaining a factor-of-20 precision improvement (from $10$ to $0.5\text{ mV}$) requires averaging $400$ independent snapshots. Measurement precision exhibits diminishing returns governed by the square root of sample size.

Level 5: intentional mistake

A student samples an $800 \text{ Hz}$ analog sine wave contaminated by high-frequency harmonic noise at $f_{\text{noise}} = 4500 \text{ Hz}$ using an ADC clocked at $f_s = 5000 \text{ Hz}$ ($5 \text{ kS/s}$). The student claims: "Since $f_s = 5000 \text{ Hz}$ exceeds my $800 \text{ Hz}$ signal by more than double, Nyquist's criterion is satisfied. I will run a digital low-pass filter in Python with cutoff at $1000 \text{ Hz}$ to remove the $4500 \text{ Hz}$ noise". Explain why this approach fails catastrophically.

Conceptual blunder: Nyquist's sampling theorem demands that sampling rate $f_s$ must exceed twice the highest frequency present in the entire analog input spectrum, including all out-of-band noise components: $f_s \geq 2 f_{\text{max}}$. Here, the Nyquist folding frequency is $f_{\text{Nyquist}} = f_s / 2 = 2500 \text{ Hz}$. Because the $4500 \text{ Hz}$ noise exceeds $2500 \text{ Hz}$, it undergoes irreversible spectral aliasing: $f_{\text{alias}} = |f_{\text{noise}} - f_s| = |4500 - 5000| = 500 \text{ Hz}$ The $4500 \text{ Hz}$ noise alias folds directly down to $500 \text{ Hz}$. Since $500 \text{ Hz}$ falls straight inside the digital filter's passband ($< 1000 \text{ Hz}$), the digital filter cannot eliminate it, permanently corrupting the target trace. Mandatory solution: Place an analog low-pass anti-aliasing filter before the ADC to eliminate $4500 \text{ Hz}$ content prior to digitization.