Travel-Time Residuals and Seismic Tomography
CSIR NET geophysics: travel-time residuals, station corrections, earthquake location, linearised tomography, damped least squares and resolution tests.
- 19 min
On this page · 14 sections
Prerequisites
- Seismology Basics and Earth’s Internal Structure — rays, ray parameter, travel-time curves and 1D Earth models.
- Seismometry — timing, picking and the instruments behind every arrival time.
- Earthquakes: Causes, Measurement and Seismicity — S–P distances and the Wadati diagram.
- Linear algebra: least squares, matrix rank, null space. The CSIR NET Part C geophysics chapter on numerical analysis and inversion develops these further.
The Hook: Arriving Early, Arriving Late
Travel-Time Residuals
Definition
$$\boxed{\delta t = t_{obs} - t_{calc}}$$where \( t_{obs} \) is the observed arrival time minus the origin time, and \( t_{calc} \) is the travel time predicted by a reference model for the catalogue location.
- Positive residual (late arrival): the ray passed through slower material than the model, on average.
- Negative residual (early arrival): faster material.
Sources of residuals
| Source | Signature |
|---|---|
| Structure near the station (crustal thickness, sediments, topography, upper-mantle anomalies) | a residual common to all events at that station: the station correction (static term); may vary with back-azimuth |
| Structure near the source (for example, a fast slab around a subduction-zone earthquake) | a residual common to all stations for that event, varying with take-off direction |
| Mislocation of the event | origin-time error: the same residual at every station; epicentre error: residuals varying as \( \cos \) of azimuth (early on one side, late on the other); depth error: residuals varying with distance |
| Deep structure along the path | residual depending on the path; this is what tomography wants |
| Picking and timing errors | random scatter; clock errors at one station |
| Ellipticity of the Earth | systematic, removed by an ellipticity correction before interpretation |
Residual from a velocity anomaly
A ray crossing a path length \( L \) in which the velocity differs from the reference \( v \) by \( \delta v \) has
$$\delta t = \frac{L}{v + \delta v} - \frac{L}{v}\qquad\approx\qquad -\frac{L}{v}\,\frac{\delta v}{v}\quad(\text{small }\delta v)$$In terms of slowness \( s = 1/v \), the relation is exactly linear in \( \delta s \) along a fixed path:
$$\boxed{\delta t = \int_{ray}\delta s\,dl}$$Earthquake Location as an Inverse Problem
Locating an earthquake means finding the hypocentre \( (x_0, y_0, z_0) \) and origin time \( t_0 \) from arrival times \( t_i \). For station \( i \),
$$t_i = t_0 + T(\mathbf x_i, \mathbf x_0)$$where \( T \) is the travel time in the model. The problem is non-linear in \( \mathbf x_0 \).
Geiger’s method (linearised iteration)
Start from a trial hypocentre and origin time. The residuals \( r_i = t_i^{obs} - t_i^{calc} \) are related to small corrections by a first-order Taylor expansion:
$$\boxed{r_i = \frac{\partial T_i}{\partial x}\Delta x + \frac{\partial T_i}{\partial y}\Delta y + \frac{\partial T_i}{\partial z}\Delta z + \Delta t_0}$$The spatial derivatives are components of the slowness vector of the ray at the source: for a station at azimuth \( \phi \) and take-off angle \( i_h \), \( \partial T/\partial x_0 = -\sin i_h\cos\phi/v_h \) (moving the source towards the station shortens the path). Four or more stations give a (usually over-determined) linear system \( \mathbf r = \mathbf G\,\Delta\mathbf m \), solved by least squares; the corrections are applied and the process is repeated until residuals stop decreasing.
What controls location quality:
- Azimuthal gap: stations on only one side leave the epicentre poorly constrained in the direction away from them.
- Depth: trades off strongly with origin time, because both shift all arrivals similarly when stations are far away. A station closer to the epicentre than about the focal depth, S arrivals, or depth phases (pP, sP) break the trade-off.
- Velocity model errors: produce systematic mislocation that formal error ellipses do not show.
Relative location methods remove much of the model error: the master-event technique and double-difference location use differences in travel times between nearby events at common stations, so path effects outside the source region cancel. They sharpen clouds of hypocentres into faults.
Joint hypocentre–velocity inversion (local earthquake tomography) solves for event locations and 3D velocity together, because each depends on the other.
Linearised Tomography
Setting up \( \mathbf d = \mathbf G\mathbf m \)
Divide the region into \( M \) cells (blocks) with unknown slowness perturbations \( \delta s_j \). For \( N \) rays, using the reference-model ray paths (justified by Fermat’s principle: first-order changes in the path do not change the travel time):
$$\boxed{\delta t_i = \sum_{j=1}^{M} l_{ij}\,\delta s_j}\qquad\text{i.e.}\qquad \mathbf d = \mathbf G\mathbf m$$where \( l_{ij} \) is the length of ray \( i \) in cell \( j \). \( \mathbf G \) is large and sparse (each ray crosses few cells).
Why it is hard
- Under-determined parts: cells crossed by no rays, or by rays that all travel in the same direction, are not constrained.
- Null space: combinations of cell values that produce zero data change. They can be added to any solution without changing the fit.
- Noise and ill-conditioning: small singular values of \( \mathbf G \) amplify noise.
Damped least squares and regularisation
The least-squares solution \( \mathbf m = (\mathbf G^T\mathbf G)^{-1}\mathbf G^T\mathbf d \) fails when \( \mathbf G^T\mathbf G \) is singular. Damped least squares (Levenberg–Marquardt, Tikhonov) minimises misfit plus a penalty on model size:
$$\boxed{\mathbf m = (\mathbf G^T\mathbf G + \varepsilon^2\mathbf I)^{-1}\mathbf G^T\mathbf d}$$- Small \( \varepsilon \): fits the data closely, but the model may be noisy and rough.
- Large \( \varepsilon \): smooth, small-amplitude model with poorer fit.
- The trade-off curve between misfit and model norm (the L-curve) helps choose \( \varepsilon \).
- Smoothing (penalising roughness, e.g. a Laplacian operator) is often used instead of, or as well as, damping.
Large systems are solved iteratively (for example with LSQR or conjugate-gradient methods) rather than by forming \( \mathbf G^T\mathbf G \). Singular value decomposition gives the cleanest view of null space and resolution for small problems.
Resolution
The model resolution matrix relates the estimated model to the true model:
$$\hat{\mathbf m} = \mathbf R\,\mathbf m_{true},\qquad \mathbf R = (\mathbf G^T\mathbf G + \varepsilon^2\mathbf I)^{-1}\mathbf G^T\mathbf G$$\( \mathbf R = \mathbf I \) means perfect resolution. Rows of \( \mathbf R \) show how each cell’s estimate is a blurred average of its neighbours. In practice, resolution is tested with synthetic models:
- Checkerboard test: compute synthetic data for an alternating pattern of fast and slow cells using the real ray geometry, add noise, invert, and see where the pattern is recovered.
- Spike or specific-structure tests: invert synthetics from a single anomaly or a hypothesised feature (for example a slab) to see how it is smeared.
- Ray-density (hit-count) maps: a necessary but not sufficient guide; many rays in one direction still resolve poorly.
A Small Tomography Problem
Four square cells of side 10 km:
| Column 1 | Column 2 | |
|---|---|---|
| Row 1 | \( s_1 \) | \( s_2 \) |
| Row 2 | \( s_3 \) | \( s_4 \) |
Rays along each row and each column (four rays, each 10 km in each of two cells):
$$d_{r1} = 10(s_1 + s_2),\quad d_{r2} = 10(s_3 + s_4),\quad d_{c1} = 10(s_1 + s_3),\quad d_{c2} = 10(s_2 + s_4)$$Four equations, four unknowns, but rank 3: \( d_{r1} + d_{r2} = d_{c1} + d_{c2} \) always. The null-space vector is the checkerboard \( (+1, -1, -1, +1) \): adding it changes no row or column sum. Row-and-column rays alone cannot see a checkerboard pattern.
Adding a diagonal ray through cells 1 and 4 (length \( 10\sqrt2 \) km in each) breaks the degeneracy, because the checkerboard adds \( +2 \) times the diagonal length to its time:
$$d_{diag} = 10\sqrt2\,(s_1 + s_4)$$This is the general lesson: crossing ray directions are what give resolution.
Tomography in Practice
Kinds of tomography
| Kind | Data | Resolves |
|---|---|---|
| Local earthquake tomography | P and S arrival times from local events, with relocations | crust and uppermost mantle beneath an active area; \( V_P \), \( V_S \), \( V_P/V_S \) |
| Teleseismic (regional) tomography (ACH-type) | relative residuals from distant events at a dense network | upper mantle beneath the network (vertical smearing) |
| Global body-wave tomography | millions of catalogue P and S times, plus differential times (e.g. PP − P, SS − S) | whole-mantle structure at long wavelengths |
| Surface-wave tomography | phase and group velocity dispersion along paths | crust and upper mantle; depth resolution from the period dependence |
| Ambient-noise tomography | cross-correlations of long noise records between station pairs give interstation surface-wave Green’s functions | crust and uppermost mantle, even without earthquakes |
| Attenuation and anisotropy tomography | amplitude decay (\( Q \)); direction-dependent velocity, shear-wave splitting | temperature, melt, fluids; flow fabric |
| Waveform and full-waveform inversion (adjoint) | complete seismograms | sharper images, at high computational cost |
Finite-frequency effects: real waves are sensitive to a volume around the geometric ray (the first Fresnel zone), not a line. For long-period body waves the sensitivity kernel is hollow along the ray itself (the “banana–doughnut” shape). Finite-frequency kernels improve imaging of small anomalies such as plume conduits.
What tomography shows
| Feature | Velocity anomaly | Interpretation |
|---|---|---|
| Subducting slabs | fast (cold) | some flatten in the transition zone, others penetrate into the lower mantle; deep slab remnants beneath former subduction zones (for example beneath India, ascribed to Tethyan subduction) |
| Mid-ocean ridges | slow in the upper ~100–200 km | hot, partially molten upwelling; shallow source |
| Cratons | fast to 200 km or more | thick, cold, depleted lithospheric keels |
| Hotspots / plumes | slow columns; some appear to extend deep | debated; resolution is the key question |
| Lowermost mantle | two large low-shear-velocity provinces (LLSVPs) beneath Africa and the Pacific; ultra-low velocity zones at the CMB | thermochemical piles; possible plume sources |
| Tibet and Himalaya | thick crust, slow mid-crust in places; fast Indian lithosphere underthrusting from the south | indentation of India |
Velocity anomalies are mostly interpreted as temperature differences, but composition, partial melt, water and anisotropy also change velocities. Comparing \( V_P \) and \( V_S \) anomalies (and their ratio) helps separate these causes.
Memory Tricks
Don’t Confuse
| This | vs | That |
|---|---|---|
| Absolute residual \( t_{obs} - t_{calc} \) | Relative residual: minus the event mean; origin-time error removed | |
| Station correction: near-receiver structure | Event correction: origin time, mislocation, near-source structure | |
| Data resolution (how well data are fitted) | Model resolution \( \mathbf R \) (how well the model is recovered) | |
| Damping: penalises model size | Smoothing: penalises roughness | |
| Ray density high | Resolution good: needs crossing directions too | |
| Absolute location (Geiger) | Relative location (double difference): path effects cancel | |
| Body-wave tomography: travel times along rays | Surface-wave tomography: dispersion; depth from period | |
| Fast anomaly: often cold (slab, craton) | Slow anomaly: hot, melt or fluids (ridge, plume) |
Common Mistakes
Practice
All questions below are original practice questions written for this page. They are not from any CSIR NET paper.
Level 1: Foundation
- Define a travel-time residual.
- What is the sign of the residual for a ray through an anomalously slow region?
- Name the test in which an alternating fast–slow pattern is used to judge resolution.
Answers
- \( \delta t = t_{obs} - t_{calc} \), observed minus predicted travel time.
- Positive (late).
- The checkerboard test.
Level 2: Part B Style
Practice Q1. A P ray travels 100 km through upper mantle with reference velocity 8.0 km/s. A region along the path is 2% slow. The residual is about:
(a) −0.26 s (b) −0.25 s (c) +0.25 s (d) +2.5 s
Answer: (c) +0.25 s
Exact: \( 100/7.84 - 100/8.0 = 12.755 - 12.500 = +0.255 \) s. Linear: \( +(100/8)\times0.02 = +0.25 \) s. Slow means late, so positive.
Practice Q2. For one earthquake, every station in a network shows a residual of about +1.2 s, regardless of distance and azimuth. The most likely cause is:
(a) a slow anomaly beneath one station (b) an origin time that is 1.2 s too early in the catalogue (c) an epicentre mislocated towards the north (d) a fast slab beneath the source
Answer: (b)
A uniform residual at all stations is the signature of origin-time error. If the catalogue origin time is too early, all observed travel times (arrival minus origin) are too long, giving positive residuals. A mislocation would give an azimuthal pattern; a station anomaly would affect one station.
Practice Q3. In damped least squares, increasing the damping parameter \( \varepsilon \) generally:
(a) improves the data fit and increases model roughness (b) reduces model amplitude and data fit, and stabilises the solution (c) has no effect on resolution (d) makes the resolution matrix closer to the identity
Answer: (b)
More damping shrinks the model towards zero (or the reference), worsens the fit, suppresses noise and moves \( \mathbf R \) further from the identity (poorer resolution).
Practice Q4. Residuals for one earthquake recorded at stations all around it are about +0.8 s at stations to the east and −0.8 s at stations to the west, varying smoothly with azimuth. The likely cause is:
(a) the true epicentre lies west of the catalogue epicentre (b) the true epicentre lies east of the catalogue epicentre (c) the origin time is too late (d) the focal depth is too shallow
Answer: (a)
East stations record late arrivals: the true source is farther from them than the catalogue location, i.e. to the west; west stations record early arrivals because the source is actually closer to them. A cosine-of-azimuth pattern means an epicentre shift.
Level 3: Part C Style
Practice Q5. In the four-cell model above (cells 10 km square), observed residuals are: row 1: +0.10 s; row 2: −0.05 s; column 1: +0.10 s; column 2: −0.05 s; diagonal through cells 1 and 4: +0.0707 s. Which statements are correct?
- The four row and column data alone cannot determine the model uniquely.
- The solution is \( s_1 = +0.010 \), \( s_2 = 0 \), \( s_3 = 0 \), \( s_4 = -0.005 \) s/km.
- Cell 1 is slower than the reference model.
- Adding the pattern \( (+0.001, -0.001, -0.001, +0.001) \) s/km to the solution leaves all five data unchanged.
(a) 1, 2 and 3 only (b) 1 and 4 only (c) 2 and 3 only (d) 1, 2, 3 and 4
Answer: (a) 1, 2 and 3 only
- (1) Rank 3 for the row/column system (checkerboard null space). True.
- (2) From rows and columns: \( s_2 = 0.01 - s_1 \), \( s_3 = 0.01 - s_1 \), \( s_4 = -0.005 - s_3 = s_1 - 0.015 \). Diagonal: \( 14.142(s_1 + s_4) = 0.0707 \) → \( 2s_1 - 0.015 = 0.005 \) → \( s_1 = 0.010 \). Then \( s_2 = s_3 = 0 \), \( s_4 = -0.005 \). True.
- (3) Positive slowness perturbation = slower. True.
- (4) The checkerboard leaves rows and columns unchanged but changes the diagonal time by \( 14.142\times0.002 = +0.028 \) s. False: that is exactly why the diagonal ray was added.
Practice Q6. A dense network above a suspected mantle plume measures relative teleseismic P residuals. Stations at the centre of the network show +0.6 s relative to the edge stations for near-vertical rays. A checkerboard test with 50 km anomalies is recovered well in map view between 100 and 300 km depth but is smeared vertically by about 100 km. Which statements are correct?
- If the anomaly occupied a vertical column 200 km tall with reference \( V_P = 8.2 \) km/s, an average slowness perturbation giving +0.6 s would correspond to about −2.4% velocity.
- The data alone could equally be fitted by a 100 km column with roughly twice the velocity anomaly.
- Relative residuals are free of origin-time errors.
- The vertical smearing means the imaged depth extent of the anomaly is uncertain by roughly 100 km.
(a) 1, 2, 3 and 4 (b) 1 and 3 only (c) 2, 3 and 4 only (d) 1 and 2 only
Answer: (a) 1, 2, 3 and 4
- (1) Linear: \( \delta v/v \approx -\delta t\,v/L = -0.6\times8.2/200 = -0.0246 \), about −2.5%; the exact value is \( v' = 200/(200/8.2 + 0.6) = 200/24.990 = 8.003 \) km/s, i.e. −2.4%. “About −2.4%” is fair. True.
- (2) With vertical rays, \( \delta t \approx -(L/v)(\delta v/v) \): halving \( L \) needs doubling \( \delta v/v \) for the same residual. This trade-off is the vertical-smearing problem. True.
- (3) Subtracting the event mean removes errors common to all stations, including origin time. True.
- (4) Consistent with the test. True.
Practice Q7. A trial location gives the following residuals (observed minus calculated, s) at four stations at equal epicentral distance, with azimuths N, E, S and W: +0.30, +0.10, −0.30, −0.10. Assume the residuals arise from an epicentre error only, and each 1 km horizontal shift towards a station reduces its travel time by 0.15 s. Which statements are correct?
- The true epicentre lies south of the trial location.
- The required shift has a south component of about 2 km and a west component of about 0.67 km.
- The mean residual is zero, so no origin-time correction is needed.
- The shift needed is about 2.1 km towards roughly S18°W.
(a) 1, 2 and 3 only (b) 1, 2, 3 and 4 (c) 2 and 4 only (d) 1 and 3 only
Answer: (b) 1, 2, 3 and 4
- N station late (+0.30), S station early (−0.30): the true source is farther from N, closer to S: south. True.
- N–S: each km south changes the N residual by −0.15 s and S by +0.15 s; removing ±0.30 needs 2.0 km south. E–W: ±0.10 needs \( 0.10/0.15 = 0.67 \) km west. True.
- Mean residual \( (0.30 + 0.10 - 0.30 - 0.10)/4 = 0 \). True.
- Magnitude \( \sqrt{2.0^2 + 0.67^2} = 2.11 \) km; direction \( \tan^{-1}(0.67/2.0) = 18.4° \) west of south. True.
Quick Revision Table
| Idea | Fact or formula |
|---|---|
| Residual | \( \delta t = t_{obs} - t_{calc} \); positive = late = slow |
| Anomaly residual | \( \delta t \approx -(L/v)(\delta v/v) \); exactly \( \int\delta s\,dl \) |
| Patterns | all stations equal → origin time; cosine of azimuth → epicentre; one station → station term |
| Relative residuals | subtract event mean; ACH teleseismic tomography |
| Geiger | \( r_i = \sum(\partial T_i/\partial x_k)\Delta x_k + \Delta t_0 \); iterate least squares |
| Location quality | azimuthal gap; depth–origin time trade-off; depth phases; model errors |
| Relative location | master event; double difference |
| Tomography | \( \delta t_i = \sum_j l_{ij}\delta s_j \); Fermat justifies fixed rays |
| Damped LSQ | \( (\mathbf G^T\mathbf G + \varepsilon^2\mathbf I)^{-1}\mathbf G^T\mathbf d \); L-curve |
| Resolution | \( \mathbf R \); checkerboard and spike tests; crossing rays needed |
| Null space example | 2 × 2 cells, rows and columns: checkerboard invisible |
| Kinds | local, teleseismic, global, surface-wave, ambient-noise, attenuation, anisotropy, full waveform |
| Images | slabs fast; ridges slow (shallow); cratons fast keels; LLSVPs under Africa and Pacific |
Key Takeaways
What’s Next
- Seismology: Quick Revision: the whole chapter on one page.
- Seismometry: timing and instrument effects behind every residual.
- Seismotectonics: Himalayan, SCR and Reservoir-Induced Earthquakes: structures that tomography images beneath India.
- Signal Processing: correlation and filtering, the tools behind picking and ambient-noise tomography.