A tensor can look stable component by component while its preferred flow direction is still wandering.
When a permeability tensor is large enough to estimate—but not yet large enough to trust
Directional upscaling gives a permeability tensor for each sampled volume. The next question is uncertainty: how much do its eigenvalues and principal directions change when the window, location or pore realisation changes?
This guide turns repeated subvolumes into intervals, angle-stability checks and a defensible representative elementary volume bound while keeping spatial dependence and near-degenerate eigenvalues visible.
Choose a reading route
Start with the route closest to the decision you need to make. Every route reaches the same worked example, transfer check and evidence limits.
- Tensor route — connect components, eigenvalues and directions.
- Bootstrap route — resample the correct spatial unit.
- Worked route — interpret value and direction intervals together.
- REV route — convert uncertainty decay into a volume decision.
- Field route — build a repeatable workflow.
Open the 12-chapter contents
- 1. A permeability tensor is estimated
- 2. Eigenvalues and directions answer different questions
- 3. Sampling windows form a distribution
- 4. Bootstrap the independent unit
- 5. Worked example: magnitudes settle before direction
- 6. Near-equal eigenvalues make directions fragile
- 7. An REV bound is a criterion
- 8. Spatial dependence widens intervals
- 9. Changed context: rotate the specimen
- 10. Bias, segmentation and boundaries
- 11. Limits and misconceptions
- 12. A practical uncertainty workflow
1. A permeability tensor is estimated
Under standard Darcy assumptions, flux relates to pressure gradient through a symmetric permeability tensor K. Matrix entries depend on coordinates; eigenvalues and eigenvectors describe principal permeabilities and directions.
One numerical solve returns one matrix. It does not show variation across neighbouring volumes or plausible reconstructions.
2. Eigenvalues and directions answer different questions
Eigenvalues measure principal magnitudes. Eigenvectors locate their axes. Magnitudes may be tight while direction is unstable, or direction stable while magnitudes remain broad.
| Diagnostic | Question | Failure signal |
|---|---|---|
| Kij interval | components stable? | wide or biased |
| eigenvalue interval | magnitudes separated? | large overlap |
| angle interval | direction repeatable? | wide cone |
| anisotropy ratio | contrast resolved? | near one with uncertainty |
3. Sampling windows form a distribution
At each candidate size L, place many admissible subvolumes across the domain. Solve the same boundary-value problem and compute K for every window.
Use the same shape and boundary convention across sizes; both can alter apparent anisotropy.
4. Bootstrap the independent unit
Voxel-by-voxel resampling destroys connectivity. Resample independent realisations, separated windows or spatial blocks large enough to preserve local structure.
Recompute the aggregate tensor and eigendecomposition in every replicate. Store values, anisotropy ratios and sign-invariant direction differences.
5. Worked example: magnitudes settle before direction
At L = 4 mm, an illustrative ensemble gives leading permeability 12.0 with 95% interval 9.0–15.5, second permeability 9.8 with 7.6–12.8, and a leading-direction cone of ±38°. Overlapping values make the axis weakly identified.
At L = 12 mm, values are 11.4 (10.8–12.1), 7.2 (6.8–7.7) and ±7°. Magnitudes and orientation now separate. This supports stability at 12 mm for the sampled process, not a universal material constant.
6. Near-equal eigenvalues make directions fragile
With equal λ1 and λ2, any orthonormal basis of their eigenspace is valid. When they are close, small tensor changes rotate eigenvectors sharply.
| Pattern | Reading | Report |
|---|---|---|
| separated values, narrow angle | resolved anisotropy | axis and cone |
| close values, wide angle | near-degenerate | plane or weak axis |
| wide values, stable angle | direction repeatable | separate magnitude claim |
| multimodal angles | multiple fabrics | show clusters |
7. An REV bound is a criterion
Predeclare tolerances, such as coefficient of variation below 10% for principal values and a 95% direction cone below 10°. The smallest L meeting every required criterion is a lower bound for that REV definition.
Porosity may stabilise before permeability because connectivity is more demanding than void fraction.
8. Spatial dependence widens intervals
Overlapping windows share voxels. Treating them as independent narrows intervals falsely. Use non-overlapping windows, regional clusters or block bootstrap with a defensible correlation length.
If the REV decision changes across plausible block sizes, report that sensitivity.
9. Changed context: rotate the specimen
Under coordinate rotation, components change, eigenvalues do not, and eigenvectors rotate with the frame. Express every tensor in a common physical frame before comparing directions.
Otherwise registration error can masquerade as material uncertainty.
10. Bias, segmentation and boundaries
Bootstrap intervals describe variability under the sampled mechanism. They do not automatically include bias from segmentation, unresolved pores, solver discretisation or boundary conditions.
Repeat the analysis across defensible thresholds and boundary conventions as separate uncertainty layers.
11. Limits and misconceptions
A narrow interval does not prove representativeness for another horizon. More windows do not repair a domain that omits large structures. REV is scale- and property-dependent.
Bootstrap percentiles can mislead with few independent units or multimodal samples; show distributions as well as endpoints.
12. A practical uncertainty workflow
Register a common frame, choose nested windows, solve identical flows, audit symmetry and positive definiteness, resample honest spatial units, match eigenvectors, plot value intervals and angle cones, then apply predeclared REV tolerances.
State whether the population is this specimen, this horizon or a broader class supported by independent samples.
Changed-context transfer check
A second soil has the same mean anisotropy ratio but bootstrap directions form two clusters 70° apart. Do not report the circular mean as a stable axis. Check for two fabrics, stratigraphic units or registration regimes and stratify or report the mixture.
A successful transfer states what stayed invariant, what changed and which conclusion must be recalculated.
Frequently asked questions
Why not bootstrap individual pores?
Permeability depends on connected paths. Independent pore resampling destroys the geometry that creates flow.
How are eigenvector sign flips handled?
Treat an eigenvector as an unoriented axis and use the smaller angle between v and ±w.
Can porosity REV replace permeability REV?
No. Connectivity makes permeability stabilise on a different scale.
What if principal eigenvalues overlap?
A unique axis may be weakly identified; report the eigenspace or direction uncertainty.
Sources and evidence trail
These sources support the definitions and limits above. The numerical teaching examples are original illustrations, not reproduced experimental results.
- Ozelim and Cavalcante on REV — micro-CT simulations and scale-dependent variability.
- Koza, Matyka and Khalili on Finite-Size Anisotropy — finite-domain anisotropy and REV.
- US Forest Service Groundwater Hydrodynamics — heterogeneity, anisotropy and REV context.
- Jones on Orientation Uncertainty — a transferable tensor-eigenvector bootstrap.
- Bootstrap Inference for Geophysical Tensors — eigenvalue and multiplicity inference.
Continue through the eduKate ecosystem
Return to Soil Science
The Gold Standard of Soil Science places this specialist problem inside the subject’s wider methods, limits and applications.
Follow the closest specialist routes
- Directional Representative Volumes, Anisotropy and Tensorial Upscaling — the estimation foundation.
- Shrink–Swell and the Reference Volume — a complementary support warning.
Continue to The Gold Standard of Education: Learning Transfers into Life and the How X Works hub.