Widen a smoothing window by a factor of six and a real 8 kilometre step in the crust can vanish from the image completely. That is not a rounding error. It is the difference between imaging two distinct crustal blocks and imaging one smooth blur that never existed. In seismic imaging, as in a lot of applied machine learning, the smoothing dial is one of the most consequential and least discussed knobs anyone touches, and the energy industry runs on maps that have all been through it.
The setup: stacking noisy picks
To image the Moho, the boundary between the crust and the mantle, receiver-function methods convert scattered seismic energy from distant earthquakes into estimates of boundary depth beneath many points along a profile, then average them. The averaging technique is common-conversion-point (CCP) stacking. Because individual picks are noisy, you do not average them equally: you weight nearby picks more heavily than distant ones using a Gaussian, the familiar bell curve. The width of that bell, sigma, is the knob. A small sigma trusts local detail; a large sigma trusts the crowd.
Why should a broad energy audience care where the Moho is? Crustal thickness sets the thermal and mechanical frame for everything above it: geothermal gradients, basin architecture, the stress state that basement faults inherit. A step in the Moho often marks a suture between two crustal blocks with different histories, and those boundaries matter for where heat, stress and fluids go. Smooth the step away and you have not just lost a line on a plot; you have merged two different geological regimes into one fictitious average.
The temptation to over-smooth
Noisy data looks unprofessional. A wide Gaussian makes it look clean, and clean plots are persuasive in a review meeting. This is exactly the trap. Smoothing does not distinguish between the noise you want gone and the signal you want kept; it attacks both. Push sigma wide enough and the profile becomes beautifully smooth and wrong. The source thesis figure that motivated this note (Fig. 3.6, a Gaussian-weighted CCP stack) is a study in getting that balance right: enough weighting to suppress scatter, not so much that genuine structure is combed flat.
Slide the dial yourself
The instrument below is the whole argument in one panel. Grey dots are noisy depth picks along a 240 km profile. The dashed line is the truth: two crustal blocks at 34 km and 42 km, separated by a real 8 km step. The aqua line is the Gaussian-weighted stack, and the amber handle is sigma.
Watch the two readouts as you widen the bell. The noise-removed number climbs quickly and early: most of the scatter is gone by a modest width while the step is still nearly intact. Keep going and the step-preserved number starts to fall, because the kernel is now reaching across the boundary and averaging the two blocks into each other. The footprint drawn under the profile shows why: at large widths the bell spans a substantial fraction of the survey, so every stacked value is a blend of both geological regimes. The right-hand panel keeps score honestly, splitting the total error into a variance term that falls with width and a bias term that rises with it. The U-shaped total has a bottom, and that bottom is the sweet spot: it does not sit at the smooth end of the dial.
Bias and variance, in the ground
Machine-learning readers will recognise this immediately as the bias-variance tradeoff. A narrow window has low bias (it faithfully follows the true depth) but high variance (it also follows the noise). A wide window has low variance (smooth, stable output) but high bias (it systematically flattens real features). No setting zeroes both; the best sigma is the one that minimises total error. The catch is that the optimum depends on how much real structure you expect and how sharp it is. A profile crossing a suture wants a narrower window than a profile over gently varying crust. The formal statement of this dilemma is decades old [4], but it is easy to forget that it applies to a stacking script just as much as to a neural network.
Why the sharp step is the hard case
A step is the worst possible target for a symmetric smoother. The Gaussian pulls values from both sides of the step toward the middle, so a true vertical jump becomes a gentle ramp. The sharper and more important the feature, the more a wide window hurts you. This is why "just smooth more" is dangerous advice in exactly the situations that matter most: abrupt boundaries are usually the geologically interesting ones, and they are the first casualties of an aggressive kernel.
Choosing sigma like an engineer
The width should be chosen by validation, not by aesthetics. Treat sigma as a hyperparameter and let the data vote: hold out a fraction of picks, sweep the width, and choose the setting that best predicts the held-out points. Where independent constraints exist, a known boundary from a nearby well or an active-source line, tune to preserve that feature rather than to minimise raw scatter. Bandwidth selection is a solved statistical problem with a mature literature [5]; the failure mode in practice is not the absence of a method but the habit of inheriting a width from last year's script.
And report the sigma you used. A Moho map without its smoothing width is like a photograph without its exposure: you cannot judge what has been averaged away. Two maps of the same region stacked at different widths can disagree about whether a boundary exists at all, and a reader who does not know the widths cannot tell which one to believe.
Zooming out
The smaller Gaussian "won" in our title not because narrow is always right, but because the discipline that makes it win generalises far beyond seismology. Any time you average noisy measurements, smoothing a production log, denoising a sensor stream, gridding sparse well data, regularising a model, you are trading detail for stability, and the setting that looks cleanest is rarely the setting that is most correct. The right instinct is not "how smooth can I make this?" but "what is the smallest smoothing that controls the noise while keeping the feature I came here to find?"
Limitations
The instrument in this note is a schematic, built to teach the tradeoff rather than to reproduce a specific stacked volume: the profile geometry, noise level and step size are illustrative, and only the smoothing behaviour follows the source thesis figure. The one-dimensional picture also simplifies real CCP practice, where the effective footprint is set jointly by the Fresnel zone of the converted waves, station spacing and depth, so sigma is not a free choice at every depth. The step-versus-ramp argument assumes the feature of interest is sharp; where the true crust varies gently, a wider kernel is genuinely optimal and the narrow setting just preserves noise. Finally, hold-out validation picks the width that predicts the data, which is not always the width that best serves a specific interpretation question; when a decision hinges on one boundary, that boundary deserves its own targeted test.
Key takeaways
- Smoothing is a tradeoff, not a cleanup: a wider Gaussian removes noise and real structure at the same time.
- Sharp features are the danger zone. A symmetric smoother turns a true step into a false ramp, and the sharper the step the worse the damage.
- This is the bias-variance dilemma in the ground: narrow means high variance, wide means high bias, and no sigma zeroes both.
- Choose sigma by validation, not aesthetics: hold out picks, sweep the width, and let predictive error pick the winner.
- Always report the smoothing width. A Moho map without its sigma hides what has been averaged away.
References
[1] Langston, C. A. (1979). Structure under Mount Rainier, Washington, inferred from teleseismic body waves. J. Geophys. Res., 84, 4749-4762. doi:10.1029/JB084iB09p04749
[2] Dueker, K. G., and Sheehan, A. F. (1997). Mantle discontinuity structure from midpoint stacks of converted P to S waves across the Yellowstone hotspot track. J. Geophys. Res., 102, 8313-8327. doi:10.1029/96JB03857
[3] Zhu, L., and Kanamori, H. (2000). Moho depth variation in southern California from teleseismic receiver functions. J. Geophys. Res., 105, 2969-2980. doi:10.1029/1999JB900322
[4] Geman, S., Bienenstock, E., and Doursat, R. (1992). Neural networks and the bias/variance dilemma. Neural Computation, 4, 1-58. doi:10.1162/neco.1992.4.1.1
[5] Silverman, B. W. (1986). Density Estimation for Statistics and Data Analysis. Chapman and Hall, London.
[6] Source: 2018 PhD thesis, receiver-function imaging of the Moho/LAB, Fig 3.6 (Gaussian-weighted CCP stack).
EarthScan builds subsurface screening that reports its own settings. If your crustal maps arrive without their smoothing width attached, we should talk.
