Find your next learning step
When two permeability eigenvalues are almost equal, their individual axes may swing under tiny perturbations even while the plane they span remains stable; the scientific claim should match that geometry.
Open the chapter contents · Soil Science root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the question, or read from Chapter 1 for the complete argument. The teaching chapters remain open; these grouped routes form the index.
Route 1 · Read the tensor
Route 2 · Work the clustered example
Route 3 · Transfer to subspaces
Route 4 · Check evidence and limits
Route 5 · Ask and continue
CHAPTER 1 OF 12 · READ THE TENSOR
1. A permeability tensor has directions and magnitudes
For anisotropic saturated flow, a symmetric positive permeability or hydraulic-conductivity tensor maps hydraulic gradient to flux. Its eigenvectors are principal directions and its eigenvalues are the corresponding principal conductivities. In those axes, the tensor is diagonal and the geometry is easier to read.
This language can tempt an analyst to treat every computed eigenvector as a stable physical direction. That is safe only when its eigenvalue is sufficiently separated from the others relative to measurement and estimation error. If two eigenvalues are close, a small perturbation can rotate their individual eigenvectors sharply.
The important correction is not to abandon the tensor. It is to report the stable object. For a separated cluster of two eigenvalues, the two-dimensional principal subspace can remain well determined even when the two axes inside it are not.
CHAPTER 2 OF 12 · READ THE TENSOR
2. Nearly isotropic is a statement about spectral gaps
An isotropic tensor has three equal eigenvalues, so every orthonormal basis is an eigenbasis. There is no unique principal direction. A nearly isotropic tensor inherits part of that ambiguity: small eigenvalue gaps make directions sensitive.
Suppose eigenvalues are 10.0, 9.9 and 2.0 in consistent units. The first two form a cluster separated from the third by a gap near 7.9, but their internal gap is only 0.1. The high-conductivity plane may be stable, while “the fastest axis” inside that plane may be fragile.
This distinction matters scientifically. A claim about a permeable plane can be supported even when a claim about a unique azimuth within that plane is not. Reporting both as equally certain exaggerates what the data identify.
CHAPTER 3 OF 12 · WORK THE CLUSTERED EXAMPLE
3. Worked perturbation: stable plane, unstable axes
Consider the diagonal teaching tensor K = diag(10.0, 9.9, 2.0). Add a small symmetric perturbation only within the upper block: E has off-diagonal entries 0.08 between axes one and two. The perturbed upper block is [[10.0,0.08],[0.08,9.9]].
Its eigenvectors rotate within the original first–second plane. For a 2×2 symmetric block, tan(2θ) = 2b/(a−d) = 1.6, so θ is about 29°. A perturbation of 0.08 produces a large axis rotation because the internal gap is only 0.1. Yet the high-conductivity plane itself has not moved at all in this constructed example.
| Quantity | Original | Perturbed interpretation |
|---|---|---|
| High eigenvalues | 10.0, 9.9 | Split and rotated within plane |
| Low eigenvalue | 2.0 | Unchanged |
| Individual high axes | Coordinate axes | Rotate about 29° |
| High-conductivity plane | Span of axes 1 and 2 | Exactly preserved |
CHAPTER 4 OF 12 · WORK THE CLUSTERED EXAMPLE
4. Principal angles measure subspace change
When a cluster is the meaningful object, compare subspaces with principal angles rather than pairing individual eigenvectors. The largest principal angle describes the greatest mismatch between the two subspaces. Projector differences provide an equivalent computational language.
The Davis–Kahan family of results bounds subspace rotation using a perturbation or residual size divided by a spectral separation from the rest of the spectrum. The relevant gap is external to the selected cluster. A small internal gap does not destabilise the cluster subspace by itself; it destabilises the chosen basis inside that subspace.
This gives a clean reporting rule: identify the cluster; state the gap that separates it from other eigenvalues; estimate perturbation or residual size; then report a subspace-angle bound. Do not quote a single-axis angle when the eigenvalue structure supports only a plane.
CHAPTER 5 OF 12 · WORK THE CLUSTERED EXAMPLE
5. Changed context: perturb the plane as well
Now add couplings of 0.10 between each high-conductivity axis and the low axis. The high plane can tilt because the perturbation connects the cluster to the separated eigenvalue. A simple norm-over-gap estimate uses perturbation size divided by the external gap, roughly 7.9 in the teaching tensor.
If a conservative perturbation norm is 0.15, the sine of the subspace angle is bounded on the order of 0.15/7.9 ≈ 0.019 under the theorem’s applicable separation conditions, corresponding to a small angle near 1.1°. The exact bound depends on the theorem formulation and chosen intervals, so the calculation is a diagnostic, not a substitute for checking hypotheses.
The transfer is the contrast: internal mixing can rotate axes by tens of degrees while external coupling tilts the plane by only about a degree. The spectrum tells you which geometric object deserves confidence.
CHAPTER 6 OF 12 · TRANSFER TO SUBSPACES
6. Measurement uncertainty belongs before interpretation
A fitted tensor comes from measurements, boundary conditions, sample orientation and an estimation procedure. Uncertainty in those inputs induces a matrix perturbation. Repeating the eigen-decomposition on bootstrap samples or plausible tensors can reveal whether eigenvalues remain clustered and whether subspaces remain stable.
| Diagnostic | Stable conclusion | Warning sign |
|---|---|---|
| Eigenvalue cluster | Cluster separated from rest | Gap overlaps uncertainty |
| Individual axis | Tight angular distribution | Axes swap or rotate widely |
| Cluster subspace | Small principal angles | Large projector variation |
| Physical evidence | Aligns with stratigraphy or fabric | Direction lacks independent support |
Do not sort eigenvectors and then average their azimuths blindly. When close eigenvalues cross, labels can swap between samples. Compare projectors or align subspaces before summarising orientation.
CHAPTER 7 OF 12 · TRANSFER TO SUBSPACES
7. Isotropy, transverse isotropy and orthotropy
Exact isotropy reduces the tensor to a scalar multiple of the identity, so direction has no hydraulic meaning. Transverse isotropy has one distinct axis and an equal pair in the perpendicular plane; the unique axis may be stable while any basis within the plane is equivalent. Orthotropy has three distinct principal values and axes.
Nearly symmetric cases should be described with uncertainty rather than forced into crisp classes. If two eigenvalues differ by less than the estimation uncertainty, treating them as a cluster may be more honest than naming two principal azimuths.
The classification must also serve the scale of the question. A laboratory specimen, field test and numerical grid cell can produce effective tensors at different scales. A direction stable at one support volume may not transfer unchanged to another.
CHAPTER 8 OF 12 · TRANSFER TO SUBSPACES
8. From tensor geometry to seepage prediction
Directional uncertainty matters because flux depends on the full tensor and hydraulic gradient. In a principal direction, flux and gradient align. In a general coordinate system, off-diagonal terms can turn the flux away from the gradient.
If the high-conductivity plane is stable but its internal axes are not, predictions dominated by in-plane versus cross-plane contrast may still be robust. Predictions that depend on a precise in-plane azimuth may not be. Sensitivity analysis should therefore rotate bases within the plausible clustered subspace and propagate that variation to flux or seepage outputs.
This is a practical transfer from linear algebra to soil science: ask which output is invariant to a basis change inside the cluster. Stable outputs can be defended even when individual eigenvectors are not.
CHAPTER 9 OF 12 · CHECK EVIDENCE AND LIMITS
9. Limits and misconceptions
A small matrix residual is not automatically a small direction error; the spectral gap mediates the conversion. Conversely, a large change in a plotted arrow does not necessarily imply a large physical change if the associated eigenvalues are nearly equal.
Symmetry and positive definiteness should be checked before applying the standard spectral interpretation. Estimated tensors can violate these properties through noise or an unsuitable inverse procedure. Any repair or regularisation changes the object and must be documented.
Principal directions do not by themselves establish geological causes. Stratigraphy, fabric, fractures, saturation and sampling design provide the independent evidence needed to interpret the tensor physically.
CHAPTER 10 OF 12 · CHECK EVIDENCE AND LIMITS
10. Evidence and source checks
Davis and Kahan’s paper, The Rotation of Eigenvectors by a Perturbation. III, develops sharp invariant-subspace bounds using angles, spectral separation and perturbation or residual information. It supports the distinction between a clustered subspace and unstable basis vectors.
Renard and colleagues’ Laboratory determination of the full permeability tensor shows why complete tensor estimation can matter and avoids assuming principal directions in advance. Alam and Farid’s hydraulic conductivity tensor of anisotropic soils connects full, non-diagonal tensors to stratigraphic orientation and seepage patterns.
Return to The Gold Standard of Soil Science for the wider system map. Continue with tensor perturbation bounds, eigenvalue gaps and principal-direction error for the single-axis case, and compare with shrink–swell and reference volume for a different measurement-boundary problem.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Frequently asked questions
Why can an eigenvector rotate when the tensor barely changes?
Because direction sensitivity scales inversely with an eigenvalue gap. Close eigenvalues make the basis inside their subspace weakly identified.
Is the eigenvalue cluster itself a problem?
No. It may indicate a physically meaningful near-symmetry. The problem is claiming more directional precision than the cluster supports.
What should I report instead of two unstable axes?
Report the clustered subspace, its principal-angle or projector uncertainty, the cluster eigenvalue range and the gap to the rest of the spectrum.
Can bootstrap eigenvectors be averaged directly?
Not reliably when axes swap or change sign. Align bases or compare subspace projectors before summarising.
Does a stable subspace guarantee stable seepage predictions?
Only for outputs insensitive to rotations within that subspace and under the model’s other assumptions. Propagate tensor uncertainty to the actual output.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Final transfer check
Given eigenvalues 6.02, 5.98 and 1.00, treat the first two as a candidate cluster. A perturbation that mixes only their eigenvectors can rotate the displayed high axes strongly while preserving their plane. A perturbation coupling that plane to the third axis should be compared with the external gap near 4.98.
Write the conclusion in two layers: “The high-conductivity plane is stable within the estimated uncertainty” and, separately, “The azimuths inside that plane are not individually resolved.” That wording preserves useful structure without pretending the tensor identifies a unique basis.
Continue through the Gold Standard of Education for the general transfer habit, or return to the verified How X Works hub.
Top · Contents · Previous · How X Works