Noise Metrics for INAs: 0.1–10 Hz and Wideband
← Back to:Instrumentation Amplifiers (INA)
This page shows how to read 0.1–10 Hz peak-to-peak noise and wideband noise density correctly, then turn them into integrated RMS noise and real, bandwidth-defined resolution. It also provides repeatable measurement rules and pass/fail criteria so noise results stay consistent across benches, boards, and production.
What “Noise Metrics” Really Mean in an INA Datasheet
INA noise is typically reported using three different lenses: 0.1–10 Hz peak-to-peak, noise density (nV/√Hz), and integrated RMS noise. These are not interchangeable “quality scores.” Each metric answers a different system question, and each becomes meaningful only when paired with a bandwidth, a measurement window, and a clear input-referred reference point.
The 3 core metrics and what each one answers
- Answers: “How much slow variation appears in a low-frequency window?”
- Best for: DC / quasi-DC sensor resolution limits.
- Pitfall: p-p depends strongly on observation time and windowing; do not compare p-p numbers without matching conditions.
- Answers: “What is the noise floor per √Hz in the frequency region of interest?”
- Best for: Computing RMS noise over a defined bandwidth (via integration / ENBW).
- Pitfall: A single headline number can hide 1/f rise or spurs; use the curve when available.
- Answers: “What is the final RMS noise inside a stated bandwidth / filter / window?”
- Best for: Mapping to resolution, minimum detectable change, and production limits.
- Pitfall: RMS without bandwidth is not actionable; RMS must be tied to ENBW or an explicit digital window.
Three questions to pick the right metric first
- What is the effective bandwidth or observation window? (Noise becomes a number only after bandwidth/window is fixed.)
- Is the target dominated by sub-10 Hz behavior or by in-band RMS? (Low-frequency limits vs wideband floor are different problems.)
- Is the goal an RMS number or a real “minimum measurable change”? (Resolution mapping requires input-referred RMS + sensor gain chain.)
Boundary rule: low-frequency noise is not drift
A low-frequency noise plot or a 0.1–10 Hz p-p number can be inflated by slow temperature gradients, bias/leakage changes, or recovery from overload. Treat drift and bias-driven slow errors as separate budget owners; keep noise metrics tied to a defined window and a stable thermal condition.
Use noise density to compute integrated RMS over the system’s noise bandwidth (ENBW). Treat RMS ↔ p-p conversion as condition-dependent (window/assumptions), not a fixed constant.
Noise Taxonomy for INAs: 1/f, White, and “Artifacts”
Before converting any datasheet number into “resolution,” it is critical to classify what dominates the spectrum. INA noise is typically a combination of 1/f noise (low-frequency rise), a white-noise floor (flat region), and artifacts (spurs or shaped bumps caused by switching, interference, or measurement chain effects). Correct classification prevents false conclusions and avoids budgets that cannot be met in hardware.
How to read a PSD plot (fast rules)
- Low-frequency slope: rising trend indicates 1/f contribution; it dominates long windows.
- Flat platform: white-noise floor; integrate this region to estimate in-band RMS.
- Narrow peaks: spurs (mains, switching ripple, coupling); handle with a mask/limit, not with RMS averaging.
- Window sensitivity: changing FFT window, sampling rate, or bandwidth changes the displayed floor via ENBW and leakage.
What the 1/f corner changes in real systems
The “corner” is the frequency where 1/f rise meets the white floor. If the effective bandwidth or observation window leans into the 1/f region, low-frequency metrics (including 0.1–10 Hz p-p) will dominate the perceived resolution. If the bandwidth is well above the corner and the front-end remains linear, the white floor dominates the integrated RMS.
Artifacts that often masquerade as “noise”
How to budget after classification
Split the “noise” budget into three owners to keep verification unambiguous:
- LF (1/f) owner: use low-frequency windows (e.g., 0.1–10 Hz p-p) under controlled thermal conditions.
- WB (white) owner: integrate density over ENBW to compute RMS in the specified bandwidth.
- Artifacts owner: define spur masks / limits (peak X relative to full-scale or target noise budget), rather than hiding peaks inside RMS.
A PSD plot can be read as three regions: 1/f rise, flat white floor, and narrow spurs. Correctly identifying the dominant region prevents mixing interference and drift-like behavior into “noise” claims.
Understanding 0.1–10 Hz Noise (Peak-to-Peak): What It Captures and What It Hides
The 0.1–10 Hz peak-to-peak number is a windowed, low-frequency view of input-referred variation. It is designed to reflect how “stable” a DC or slow sensor reading looks over time. Unlike RMS metrics, peak-to-peak is not a stationary statistic—it changes with observation time, filtering method, and sample count. A meaningful p-p requirement must always be tied to a stated measurement window and band-pass method.
What 0.1–10 Hz p-p is actually measuring
- Step 1: Record a time series for an observation time Tobs.
- Step 2: Apply a 0.1–10 Hz band-pass (analog or digital).
- Step 3: Compute peak-to-peak as max − min over the filtered record.
- Interpretation: The result is a window-limited “how much it moves” metric, not a bandwidth-agnostic noise floor.
Why peak-to-peak is statistically unstable
Peak-to-peak is an extreme-value statistic. With more samples (longer Tobs or higher sample rate), the probability of observing a larger excursion increases. Band-pass details (transition band, group delay, digital windowing) also change the “shape” of the filtered waveform and therefore the peaks that are seen.
Common mistakes that create “inconsistent” results
- Comparing p-p to RMS directly: p-p depends on window and assumptions; RMS depends on bandwidth/ENBW.
- Comparing different record lengths: longer Tobs usually increases p-p even when the device is unchanged.
- Including slow components: thermal gradients, bias/leakage changes, or recovery effects can inflate the in-window variation.
- Including narrowband interference: mains or coupling spurs are not “noise floor” and should be handled as separate limits.
How to specify a comparable 0.1–10 Hz p-p requirement
A production-ready p-p requirement must include the minimum reporting fields below. Without them, pass/fail cannot be reproduced.
- Band-pass: 0.1–10 Hz (method: analog/digital)
- Observation time: Tobs = ___
- Sampling rate (if digital): fs = ___
- Input condition: shorted / stated source impedance
- Thermal condition: steady-state (no airflow steps)
Peak-to-peak is taken after the 0.1–10 Hz band-pass and within the chosen observation window. Longer windows increase the chance of larger extremes.
Wideband Noise Density (nV/√Hz): How to Read and Use It
Noise density expresses the RMS noise per √Hz at a given frequency. It becomes a system number only after integration over the effective noise bandwidth set by the analog filter or the digital window. Because INA density is often frequency-dependent (1/f rise at low frequency and a flat white floor at higher frequency), a single headline value can be misleading unless it matches the frequency region of interest.
What nV/√Hz means (system translation)
- Density domain: noise is expressed per √Hz at a specific frequency.
- RMS domain: RMS noise is obtained by integrating density over ENBW.
- Design implication: increasing bandwidth increases integrated RMS even if the density curve is unchanged.
How to read the curve (three read points)
Use the curve to identify whether the operating band sits in the 1/f rise or in the white floor. A single headline density value is only valid when it is read in the same region used for integration.
- 10 Hz: checks low-frequency rise proximity.
- White floor: the flat region used for most wideband RMS calculations.
- 1 kHz: a practical in-band anchor for many conditioned sensors.
Source resistance noise (combine in density first)
A sensor or source resistance contributes its own noise in the same nV/√Hz domain. Combine independent density contributors using an RMS-in-density approach before integrating:
Production-ready reporting fields for density
- Density point: en @ f = ___ nV/√Hz
- Gain and input reference: input-referred statement
- Source impedance stated (shorted or Rs = ___)
- Bandwidth / filter context (for later integration)
- Spur handling: spur mask / limits kept separate
A typical INA density curve rises at low frequency (1/f) and flattens into a white floor. Read density at representative points, then integrate over ENBW to obtain RMS.
Converting Noise Density to Integrated RMS Noise
Integrated RMS noise is the input-referred RMS voltage that results after the noise density curve is shaped by a measurement bandwidth or a filter/window. This number is only meaningful when the noise bandwidth is stated. In practice, the fastest path is to combine density contributors in the nV/√Hz domain, use ENBW to represent the filter/window, then integrate to obtain Vrms and translate it through gain to the output domain.
Model 1: Pure white noise (fast engineering estimate)
Use when the operating band sits in the white floor region of the density curve (away from the 1/f rise). Read a representative en_white, then apply the noise bandwidth.
- Read en_white from the flat region of en(f).
- Use BW_noise or ENBW as the bandwidth.
- Compute Vrms_in ≈ en_white · √(BW_noise).
- Translate: Vrms_out = Vrms_in · Gain.
Model 2: 1/f + white (corner-based, piecewise workflow)
Use when low-frequency rise contributes to the band. Treat the curve as two regions separated by a corner frequency where the 1/f rise meets the white floor. This avoids academic math while preserving the correct dependency on the operating limits.
- Find f_c from en(f): where the 1/f rise meets the flat floor.
- Define band limits f_L and f_H from the measurement window/filter.
- Estimate low-frequency contribution using the curve segment over [f_L, f_c].
- Estimate white-floor contribution using en_white over [f_c, f_H].
- Combine the two contributions in RMS (do not add amplitudes).
Model 3: With a filter/window (use ENBW, not -3 dB)
Analog anti-alias filters, RC networks, and digital averaging windows shape noise. Represent that shaping using ENBW (effective noise bandwidth). Once ENBW is known, apply Model 1 or Model 2 using BW_noise = ENBW.
- Filter/window type and settings
- ENBW value (from datasheet, sim, or measurement)
- Band limits f_L, f_H (if applicable)
- Gain and reference point (input-referred statement)
Sanity checks (catch wrong bandwidth and spur mixing)
Combine independent contributors in the density domain, apply the filter/window through ENBW, integrate to input-referred Vrms, then translate through gain.
RMS ↔ Peak-to-Peak: When the Conversion is Valid (and When It’s Not)
The ratio k = Vpp / Vrms is not a universal constant. Peak-to-peak is an extreme-value statistic that depends on observation time, bandwidth/window, and whether the record contains slow components that are not part of broadband noise. A conversion is only meaningful when the signal is approximately Gaussian and the measurement window and bandwidth are fixed and explicitly stated.
When a conversion is valid
- Noise is close to Gaussian within the stated band.
- Window and bandwidth are fixed: same Tobs and same band-pass/ENBW.
- Measurement conditions are steady: no step changes in temperature, recovery, or bias state.
- Spurs are treated separately: the record is not dominated by narrowband interference.
Why “one k” fails across different windows
Peak-to-peak grows with longer observation because more samples increase the chance of seeing a larger extreme. RMS does not grow the same way when bandwidth is fixed. Therefore, k changes with Tobs even when the underlying density curve is unchanged.
How to self-calibrate k in a real system (recommended)
- Fix the measurement definition: Tobs and the band-pass/ENBW.
- Fix the input condition: shorted input or stated source impedance.
- Capture one record and compute Vrms and Vpp on the same window.
- Repeat across multiple records to obtain a k distribution.
- Pick a conservative k rule that matches the acceptance philosophy (range-based).
- Store k together with the test definition as part of the production spec.
Boundary reminder: do not mix slow components into noise conversion
If the record contains slow components (thermal gradients, bias/leakage changes, overload recovery), Vpp inflates and k no longer represents broadband noise. Treat slow error owners under DC accuracy and keep noise conversion limited to a defined band/window.
For a fixed bandwidth, longer observation usually increases the extreme values seen in a record. That pushes peak-to-peak upward and increases k.
Converting Noise Density to Integrated RMS Noise
Integrated RMS noise is the input-referred RMS voltage that results after the noise density curve is shaped by a measurement bandwidth or a filter/window. This number is only meaningful when the noise bandwidth is stated. In practice, the fastest path is to combine density contributors in the nV/√Hz domain, use ENBW to represent the filter/window, then integrate to obtain Vrms and translate it through gain to the output domain.
Model 1: Pure white noise (fast engineering estimate)
Use when the operating band sits in the white floor region of the density curve (away from the 1/f rise). Read a representative en_white, then apply the noise bandwidth.
- Read en_white from the flat region of en(f).
- Use BW_noise or ENBW as the bandwidth.
- Compute Vrms_in ≈ en_white · √(BW_noise).
- Translate: Vrms_out = Vrms_in · Gain.
Model 2: 1/f + white (corner-based, piecewise workflow)
Use when low-frequency rise contributes to the band. Treat the curve as two regions separated by a corner frequency where the 1/f rise meets the white floor. This avoids academic math while preserving the correct dependency on the operating limits.
- Find f_c from en(f): where the 1/f rise meets the flat floor.
- Define band limits f_L and f_H from the measurement window/filter.
- Estimate low-frequency contribution using the curve segment over [f_L, f_c].
- Estimate white-floor contribution using en_white over [f_c, f_H].
- Combine the two contributions in RMS (do not add amplitudes).
Model 3: With a filter/window (use ENBW, not -3 dB)
Analog anti-alias filters, RC networks, and digital averaging windows shape noise. Represent that shaping using ENBW (effective noise bandwidth). Once ENBW is known, apply Model 1 or Model 2 using BW_noise = ENBW.
- Filter/window type and settings
- ENBW value (from datasheet, sim, or measurement)
- Band limits f_L, f_H (if applicable)
- Gain and reference point (input-referred statement)
Sanity checks (catch wrong bandwidth and spur mixing)
Combine independent contributors in the density domain, apply the filter/window through ENBW, integrate to input-referred Vrms, then translate through gain.
RMS ↔ Peak-to-Peak: When the Conversion is Valid (and When It’s Not)
The ratio k = Vpp / Vrms is not a universal constant. Peak-to-peak is an extreme-value statistic that depends on observation time, bandwidth/window, and whether the record contains slow components that are not part of broadband noise. A conversion is only meaningful when the signal is approximately Gaussian and the measurement window and bandwidth are fixed and explicitly stated.
When a conversion is valid
- Noise is close to Gaussian within the stated band.
- Window and bandwidth are fixed: same Tobs and same band-pass/ENBW.
- Measurement conditions are steady: no step changes in temperature, recovery, or bias state.
- Spurs are treated separately: the record is not dominated by narrowband interference.
Why “one k” fails across different windows
Peak-to-peak grows with longer observation because more samples increase the chance of seeing a larger extreme. RMS does not grow the same way when bandwidth is fixed. Therefore, k changes with Tobs even when the underlying density curve is unchanged.
How to self-calibrate k in a real system (recommended)
- Fix the measurement definition: Tobs and the band-pass/ENBW.
- Fix the input condition: shorted input or stated source impedance.
- Capture one record and compute Vrms and Vpp on the same window.
- Repeat across multiple records to obtain a k distribution.
- Pick a conservative k rule that matches the acceptance philosophy (range-based).
- Store k together with the test definition as part of the production spec.
Boundary reminder: do not mix slow components into noise conversion
If the record contains slow components (thermal gradients, bias/leakage changes, overload recovery), Vpp inflates and k no longer represents broadband noise. Treat slow error owners under DC accuracy and keep noise conversion limited to a defined band/window.
For a fixed bandwidth, longer observation usually increases the extreme values seen in a record. That pushes peak-to-peak upward and increases k.
Mapping Noise to Real Sensor Resolution and Bandwidth
Noise metrics become actionable only after converting them into a minimum detectable change under a stated bandwidth and observation definition. Start with input-referred integrated noise (Vrms_in over BW/ENBW), translate it through the signal chain and sensor sensitivity, then express the result as Δsensor_min or an LSB-equivalent value. Bandwidth is the main lever: for broadband noise, larger noise bandwidth increases Vrms and reduces achievable resolution.
Step 1: Use a single reference domain (input-referred)
Resolution mapping is most robust when all noise is expressed as input-referred Vrms. If a measurement is made at the output or ADC input, translate back using the stated gain. Keep the bandwidth definition attached to the number.
- Vrms_in over BW/ENBW
- Total gain to the measurement point
- Sensor sensitivity S (V per unit or equivalent)
- Observation definition (window / method)
Step 2: Convert Vrms into a minimum detectable input change
A detection threshold must be defined (decision rule, averaging, and bandwidth). Use a simple template that preserves the dependency on bandwidth without injecting application-specific constants.
Step 3: Optional LSB-equivalent mapping (measurement chain)
When a digital limit is needed, express the integrated noise at the converter input as a fraction of a code step. Keep the mapping generic and avoid mixing it with converter architecture details.
Bandwidth–resolution tradeoff (and a fast validation)
For broadband noise, integrated RMS increases as noise bandwidth increases. That reduces achievable resolution unless averaging or bandwidth limiting is applied. Validate the noise owner by changing BW/ENBW and checking whether Vrms follows the expected scaling.
Spec writing pattern (make resolution reproducible)
- Δsensor_min reported with BW/ENBW = ___
- Observation definition: window / method = ___
- Input condition: source impedance / shorted = ___
- Gain mapping and sensitivity fields stated
Use a generic chain: sensitivity and gain translate integrated input noise into the minimum detectable sensor change. Bandwidth affects Vrms directly and therefore sets the resolution limit.
Chopper / Zero-Drift INAs: Noise Ripple, Folding, and How to Budget It
Zero-drift and chopper INAs often deliver excellent low-frequency stability, but they can introduce structured artifacts that do not behave like random broadband noise. Typical signatures include a fixed spur in the spectrum or a periodic ripple in the time domain. A production-ready noise budget must separate broadband Vrms from spur/ripple limits and evaluate whether the measurement bandwidth window covers the artifact.
Noise viewpoint (benefit vs cost)
- Benefit: improved low-frequency stability and reduced low-frequency noise.
- Cost: ripple/spur signatures that can land inside the measurement band.
- Risk: artifacts can dominate “resolution” even when the density floor is low.
Separate two acceptance limits
Bandwidth window decision (in-band vs out-of-band)
The artifact becomes a resolution limiter when the measurement BW window covers it. If the window excludes the spur, verify that the filter/window provides enough attenuation and that the spur does not fold back through sampling or processing.
- In-band spur: treat as a ripple/periodic error; set a peak limit.
- Out-of-band spur: require attenuation margin; verify with spectrum.
Handling principles (keep topology on the AAF page)
Measurable pass criteria (separate Vrms and spur limits)
A fixed spur is a structured signature. The key question is whether the measurement bandwidth window covers it. In-band spurs must be handled as ripple or spur limits; out-of-band spurs require verified attenuation margin.
Measurement Setups: How to Measure 0.1–10 Hz and Wideband Noise Correctly
Accurate noise measurement depends on a fixed definition: bandwidth (BW/ENBW), observation time, input condition, and processing method. Low-frequency noise is dominated by test environment and stability (shielding, thermal equilibrium, and long observation), while wideband noise demands bandwidth control, FFT/ENBW reporting, and a quantified instrument noise floor. Without measurement-chain self-noise characterization, results cannot be compared to datasheet numbers.
Low-frequency (0.1–10 Hz) setup: stabilize first, then measure
- Shielding: use an enclosure and keep cables fixed.
- Thermal equilibrium: wait for steady conditions; avoid airflow changes.
- Tobs: state observation time; peak-to-peak depends on it.
- Band-pass definition: specify analog or digital implementation.
- Input condition: report shorted input vs stated source impedance.
Wideband setup: control bandwidth, then report FFT with ENBW
- Front-end BW: define the measurement bandwidth or anti-alias limit.
- Instrument floor: quantify ADC/FFT self-noise before trusting results.
- FFT settings: report window type and correction approach.
- ENBW: use ENBW/RBW consistent reporting for comparability.
Quantify measurement-chain self-noise (required)
If the measurement chain is not substantially quieter than the DUT, the reported noise is dominated by the instrument and the conclusion is not actionable.
- Measure input short baseline (chain floor).
- Measure with a known source (sanity verification).
- Change gain and verify noise scaling trend.
- Use RMS power separation only when chain floor is well below the measurement.
Fast sanity checks (detect definition drift)
Reporting template (copy-ready fields)
A stable chain separates DUT noise from instrument floor and ties every number to bandwidth and observation definitions.
Common Traps: When “Noise” Is Actually Drift, Hum, Leakage, or Saturation Recovery
A large fraction of “noise problems” are not random broadband noise. The fastest way to converge is to classify the signature and apply a short diagnostic test. The traps below provide a quick check, a bounded fix action, and a pass-criteria placeholder that keeps noise reporting reproducible without expanding into layout or EMI theory.
Trap A: 50/60 Hz hum (and harmonics)
Trap B: thermal gradients / airflow sensitivity
Trap C: leakage / contamination / bias path issues
Trap D: saturation / overload recovery
Use a fast classification pass to avoid integrating spurs or slow components into a broadband noise metric.
Engineering Checklist: Noise Budgeting, Verification, and Production Screens
This checklist turns noise metrics into a reusable workflow: define a minimal noise budget, verify with reproducible measurements, then deploy production screens that are short-time, automatable, and traceable. The scope is noise-only (bandwidth/ENBW, observation time, floor, spurs, repeatability) and does not mix in CMRR, offset/drift, or protection topics.
Design checklist (minimal noise budget fields)
- en(f) curve source and region of use
- 0.1–10 Hz p-p and stated conditions
- Noise bandwidth: BW/ENBW (not -3 dB BW)
- INA gain and tolerance
- Sensor sensitivity S (placeholder field)
- Target resolution: Δsensor or LSB-equivalent
- Margin placeholder (X% or X dB)
- Noise owner assignment (DUT / ADC / BW / environment)
- Verification plan tied to budget numbers
Verification checklist (five must-measure items)
Production screens (short-time + automatable + traceable)
Production cannot afford long observation for low-frequency peak-to-peak. Use short-time RMS and spur detection with a fixed bandwidth definition, plus a temperature audit strategy per lot or shift.
Reference BOM (starting points only)
These part numbers are provided to speed up fixture and measurement-chain setup. Final selection depends on availability, packaging, and the required bandwidth/precision definition.
Use the left checklist to enforce consistent definitions, and the right table to standardize recording fields across design, verification, and production.
Request a Quote
FAQs: Noise Metrics (0.1–10 Hz & Wideband) — Practical Measurement and Mapping
These FAQs close common long-tail questions about noise specifications, measurement definitions, and mapping noise into resolution. Each answer follows a fixed 4-line, data-oriented structure.
