How certain is
the wave estimate?
A spectrum is an estimate of a changing physical process, not an error-free picture of the sea. Separate sampling variability, instrument and processing uncertainty, and natural sea-state variability before making engineering decisions.
30.1 · What is uncertain?
Let η(t) denote sea-surface elevation (m), with one-sided variance-density spectrum Sη(f) (m² Hz⁻¹), for positive frequency f (Hz). A measured record x(t) may be modelled schematically as x(t)=g[η(t)]+b(t)+ε(t), where g represents the sensor and its response, b a systematic contribution, and ε random measurement noise. In practice, the observation model may be frequency dependent, nonlinear and affected by orientation, depth, drift and environmental conditions.
Type A evaluations use statistical analysis of observations; Type B evaluations use other information such as calibration certificates, manufacturer specifications, prior characterization and physical models. These classifications describe the evaluation method, not a simple division into random and systematic errors. A known significant bias should be corrected where justified, with uncertainty retained for the correction.
30.2 · Distribution and confidence interval of a spectral estimator
For a locally stationary, approximately Gaussian process, suitably normalized independent periodogram ordinates are asymptotically exponentially distributed. Averaging independent estimates gives the useful approximation ν Ŝ(f)/S(f) ∼ χ²ν, with ν effective degrees of freedom (DOF). For K independent, equally weighted, unsmoothed periodograms, ν≈2K. This is a sampling model, not an exact finite-record guarantee.
Here Ŝ is the estimated one-sided spectral density (m² Hz⁻¹), χ²p,ν is the p-quantile of the chi-square distribution, α is the two-sided error probability, and ν is the estimator's justified effective DOF. The interval is asymmetric because χ² is skewed. The commonly quoted relative standard deviation is approximately √(2/ν), valid under this estimator model.
Interactive experiment A · Sampling confidence interval
Observe how more independent information narrows the interval. The displayed interval is pointwise and conditional on a stationary Gaussian spectral-estimation model.
30.3 · Welch averaging, leakage and effective information
Welch's method divides N samples (sample interval Δt in s) into K segments of L samples, applies a window w[n], computes modified periodograms and averages them. With hop D samples, K=1+⌊(N−L)/D⌋ when the record is fully populated. The nominal Fourier spacing is Δf=fs/L (Hz), with fs=1/Δt. Window main-lobe width and equivalent noise bandwidth (ENBW) affect practical resolving power: ENBW = fs Σw[n]²/(Σw[n])², in Hz. Zero-padding interpolates displayed frequencies; it does not create independent resolution.
Overlapping, windowed segments are generally correlated. An approximate Gaussian-process correction is νeff≈2K/[1+2 Σj=1…K−1(1−j/K)ρj²], where ρj=Σnw[n]w[n+jD]/Σnw[n]² (zero outside the window). This approximation addresses overlap covariance for a given window and is not universally valid for all spectral smoothing, processes or frequencies. More general weighted averages require a covariance-based DOF estimate; do not infer ν from K alone.
Interactive experiment B · Resolution versus stability
Illustration assumes N=4096 samples, fₛ=2 Hz, Hann window, and the stated approximate overlap-DOF model. Greater overlap does not produce proportionally more independent estimates.
30.4 · Pointwise intervals are not simultaneous bands
A 95% pointwise interval has approximately 95% coverage at one preselected frequency. Plotting 95% intervals at many frequencies does not yield a 95% probability that the entire true spectrum lies inside all intervals. For M independent tests, the probability of simultaneous coverage would be 0.95M; correlated spectral bins invalidate this simple product.
A conservative simultaneous construction applies Bonferroni: use individual two-sided significance α/M at each of M predeclared frequencies, yielding familywise coverage at least 1−α under valid marginal intervals, regardless of dependence. This may be wide. Resampling, parametric simulation or covariance-aware methods can provide less conservative bands if the data-generating model and dependence are defensible. Frequency selection after inspecting the data needs special care.
Interactive experiment C · Multiple-frequency coverage
30.5 · Spectral moments and derived wave parameters
For integration bounds fa and fb (Hz), define mn=∫fₐfᵦ fnSη(f)df, with units m² s−n. In discrete form, m̂n=ΣifinŜiΔfi. Thus Hm0=4√m0 (m), Tm01=m0/m1 (s), and Tm02=√(m0/m2) (s). The factor four defines the conventional spectral significant wave height; it is not an assertion that every sea-state distribution is Rayleigh.
Write Cij=Cov(Ŝi,Ŝj) (m⁴ Hz⁻²). Then Cov(m̂p,m̂q)=Σij(fipΔfi)(fjqΔfj)Cij. Omitting off-diagonal terms is justified only if their contribution is negligible. For g(m), the first-order delta method gives u²(g)≈∇gᵀCm∇g. For Hm0, ∂H/∂m0=2/√m0; therefore u(H)≈(2/√m0)u(m0) and relative uncertainty u(H)/H≈½u(m0)/m0.
For Tm01, the gradient with respect to (m0,m1) is (1/m1, −m0/m1²). For Tm02, the gradient with respect to (m0,m2) is (Tm02/(2m0), −Tm02/(2m2)). Their uncertainties must include moment covariance. The peak period Tp=1/fp is especially unstable for flat or multi-peaked spectra; a first-order derivative cannot capture peak switching or bin-selection effects. Use realizations, spectral peak fitting with validated assumptions, or a Monte Carlo analysis where appropriate.
Interactive experiment D · Propagating uncertainty into Hm0
First-order propagation only; a ±u range is not automatically a 68% confidence interval or an expanded measurement uncertainty.
30.6 · Traceable measurement uncertainty budget
Build a measurement model from the actual sensor, acquisition chain, response correction and spectral-processing pipeline. Potential contributions include calibration scale and offset; sensor frequency response and transfer-function inversion; tilt/orientation and pressure-to-surface conversion; clock drift, synchronization and sampling jitter; digitization, resolution and clipping; mooring motion, depth and atmospheric pressure; acoustic contamination; missing data and QC choices; window leakage, finite bandwidth and high-frequency truncation. Avoid counting the same noise source twice when a statistical repeatability estimate already includes it.
| Contribution | Standard uncertainty (m) | Evaluation / dependency |
|---|---|---|
| Repeatability of Hm₀ estimator | 0.09 | Type A; conditional on stationary record |
| Calibration and sensor response | 0.06 | Type B; response characterization |
| Processing and bandwidth | 0.04 | Type B or A; sensitivity experiments |
| Environmental correction | 0.03 | Type B; site-specific model |
For this illustrative independent-component budget, uc=√(0.09²+0.06²+0.04²+0.03²)=0.119 m. For Hm0=3.00 m, a nominal k=2 expanded uncertainty is U≈0.238 m. Coverage probability cannot be asserted as exactly 95% without distributional and effective-DOF justification. Correlated terms require uc²=Σui²+2Σi<jrijuiuj, with rij correlation coefficients.
Monte Carlo propagation: define joint probability distributions for uncertain inputs, including correlations; sample input vectors; pass each sample through the same measurement and spectral-processing model; summarize the resulting output distribution and its quantiles. Verify convergence and sensitivity, document distribution choices and distinguish propagated input uncertainty from stochastic variability of the sea state. A time-series bootstrap must respect serial dependence (e.g. blocks), and may fail under nonstationarity.
30.7 · Nonstationarity, decisions and reporting
A storm ramp, evolving swell system, tidal modulation or abrupt sensor change violates the constant-spectrum premise over a record. Splitting a nonstationary record into shorter windows may reduce bias from changing conditions but increases estimator variability and worsens frequency resolution. Diagnostics include time-frequency displays, segment-by-segment moments, changes in variance and spectral shape, and contextual environmental records. A stationary χ² interval must not be used as a confidence interval for a changing physical spectrum without qualification.
Report the sampling rate and duration; sensor and calibration provenance; instrument response and uncertainty corrections; detrending and QC flags; spectral window, segment length, overlap, ENBW, normalization and integration limits; estimated DOF and its method; confidence level and whether pointwise or simultaneous; spectral moment definitions; treatment of covariance and nonstationarity; and any Type A/Type B budget, coverage factor and assumptions. For design or model validation, distinguish instrument uncertainty, sampling uncertainty, spatial representativeness, model discrepancy and longer-term climate variability.
30.8 · Methods and reference provenance
Metrology guidance: JCGM 100:2008, Evaluation of measurement data — Guide to the expression of uncertainty in measurement (GUM); JCGM 101:2008, Supplement 1 — Propagation of distributions using a Monte Carlo method. These govern uncertainty evaluation concepts, not wave-specific numerical thresholds. Spectral methodology: Welch, P. D. (1967), “The use of fast Fourier transform for the estimation of power spectra: A method based on time averaging over short, modified periodograms,” IEEE Transactions on Audio and Electroacoustics 15(2), 70–73, DOI 10.1109/TAU.1967.1161901; Percival, D. B. & Walden, A. T. (1993), Spectral Analysis for Physical Applications, Cambridge University Press. Wave science: Holthuijsen, L. H. (2007), Waves in Oceanic and Coastal Waters, Cambridge University Press, especially the spectral-analysis appendix; Kamphuis, J. W. (2000), Introduction to Coastal Engineering and Management, World Scientific, short-term wave analysis. Data context: IOOS QARTOD Manual for Real-Time Quality Control of In-Situ Surface Wave Data, version 2.1 (2019), DOI 10.25923/7yc5-vs69; CF Conventions 1.10 for exchange metadata. These sources have distinct roles; no claim of compliance or complete clause-by-clause review is made here.