Find your next learning step
Replicate eigenspaces cannot be averaged by averaging their basis vectors. A defensible consensus treats each subspace as a geometric object, keeps dispersion visible and tests whether one mean is even representative.
Open the chapter contents · Polymer 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 chapters remain open on the page; these grouped routes are an index.
Route 1 · Define the estimand
Route 2 · Work the replicate ledger
Route 3 · Test recovery and spread
Route 4 · Check limits and evidence
Route 5 · Ask and continue
CHAPTER 1 OF 12 · DEFINE THE ESTIMAND
1. Replicates return subspaces, not matching arrows
Suppose each polymer replicate yields a symmetric conformation tensor with a nearly repeated pair of eigenvalues. The stable object is a two-dimensional eigenspace. Each software routine may return a different orthonormal basis for that same plane, so averaging columns as ordinary vectors can cancel or invent directions.
Represent replicate i by an orthonormal matrix U_i and projector P_i=U_iU_i^T. Replacing U_i by U_iR for any orthogonal within-plane rotation R leaves P_i unchanged. This removes sign flips and arbitrary basis rotations before averaging begins.
CHAPTER 2 OF 12 · DEFINE THE ESTIMAND
2. Define the mean and the dispersion together
One extrinsic consensus minimises squared Frobenius distances between projectors. Compute the weighted matrix A=Σw_iP_i, then take the r-dimensional eigenspace associated with its largest eigenvalues. This is often called a projection or chordal mean.
An intrinsic Fréchet mean instead minimises squared geodesic distances on the Grassmann manifold. The two means need not coincide. State the metric, weights, dimension and optimisation method, then report dispersion; one consensus without spread can hide a split population.
CHAPTER 3 OF 12 · WORK THE REPLICATE LEDGER
3. Work a three-replicate projector ledger
Let the target be a line in the plane so the arithmetic is visible. Replicate directions are at 0°, 10° and 20°. Their one-dimensional projectors ignore the sign of each direction. Averaging the three projectors gives a leading eigenvector at 10°, the symmetric consensus.
| Replicate | Direction angle | Angle to consensus | Projector interpretation |
|---|---|---|---|
| A | 0° | 10° | lower-side replicate |
| B | 10° | 0° | central replicate |
| C | 20° | 10° | upper-side replicate |
For a clustered plane in three dimensions, the same calculation uses rank-two projectors. Principal angles compare each replicate plane with the consensus without aligning arbitrary in-plane axes.
CHAPTER 4 OF 12 · WORK THE REPLICATE LEDGER
4. Read the spectrum of the average projector
Every P_i has eigenvalues zero or one, but A generally has values between them. Large leading eigenvalues mean the corresponding directions appear consistently across replicates. A wide gap between eigenvalues r and r+1 supports a stable r-dimensional consensus under the chosen weights.
If the gap is small, the estimated mean subspace is sensitive. Bootstrap the entire replicate process and report principal-angle intervals. Do not interpret a smooth-looking mean basis when A itself says the dimension boundary is uncertain.
CHAPTER 5 OF 12 · WORK THE REPLICATE LEDGER
5. Compare consensus recovery across time
At each recovery time, estimate the consensus projector and its replicate spread. Compare the time consensus with the pre-shear reference by principal angles or projector distance. Then separately compare dispersion: a mean may return before the replicates become mutually consistent.
| Recovery pattern | Mean distance to baseline | Replicate spread | Reading |
|---|---|---|---|
| early | large | large | shifted and heterogeneous |
| middle | small | large | average recovered, replicates disagree |
| late | small | small | consensus and reproducibility recovered |
CHAPTER 6 OF 12 · TEST RECOVERY AND SPREAD
6. Changed context: one mean hides two modes
Now place half the replicate lines at 0° and half at 90°. The average projector is one half times the identity. Every direction is a leading direction, so no unique one-dimensional projection mean exists. A numerical routine may return an arbitrary axis and falsely present it as a consensus.
This transfer check distinguishes broad noise from genuine multimodality. Inspect pairwise principal angles, cluster the subspaces when scientifically justified, and report both modes. A single mean is not mandatory when the sample geometry rejects it.
CHAPTER 7 OF 12 · TEST RECOVERY AND SPREAD
7. Choose weights before seeing the preferred answer
Equal weights treat replicates as exchangeable. Precision weights may be justified when tensor uncertainty differs, but weights estimated from the same data can amplify selection bias. A long trajectory sampled many times should not automatically outweigh independent specimens.
Declare the experimental unit. Use hierarchical resampling when several frames belong to one specimen. Test sensitivity to equal, specimen-balanced and uncertainty-based weights. The biological or material population, not convenient file count, determines the averaging unit.
CHAPTER 8 OF 12 · TEST RECOVERY AND SPREAD
8. Connect the geometry to polymer evidence
A consensus eigenspace describes a tensor-level orientation pattern. It does not identify a unique molecular configuration. Compare it with eigenvalues, trace, determinant, stress response and the protocol that created the tensor.
Use Spectral Projectors, Clustered Eigenspaces and Subspace-Angle Recovery for the single-replicate geometry and Eigenvalue Degeneracy, Axis Uncertainty and Frame-Invariant Recovery Measures for deciding when individual axes are not identifiable.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Limits and common misconceptions
The Euclidean mean of projectors is not itself usually a projector; the eigenspace extraction is essential. An intrinsic mean can have local minima or cease to be unique when observations spread too widely. A tight mean does not correct biased measurement or a wrong cluster dimension.
Projector averaging assumes a fixed subspace dimension unless the method explicitly models changes. Temporal smoothing reduces apparent spread and can manufacture recovery. A bootstrap over frames is not a bootstrap over specimens when frames share the same material history.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Build an evidence check
Archive each tensor, eigenvalues, cluster rule, U_i, P_i, specimen identity, time, uncertainty and preprocessing. Recompute the mean under the declared metric and weights. Report the average-projector spectrum, pairwise principal angles, distance to baseline and bootstrap intervals.
Run sign flips and within-subspace rotations as invariance tests. Add a synthetic split-mode case and verify that the procedure signals non-uniqueness rather than returning a confident axis. Test whether the summary predicts held-out directional stress better than simpler tensor invariants.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Questions readers should ask
Why not average eigenvectors directly?
Their signs and bases are arbitrary. Projectors represent the subspaces without those coordinate choices.
Is a projection mean the same as an intrinsic Grassmann mean?
No. They minimise different distances and should be named with the metric used.
Can the mean recover while replicates remain unstable?
Yes. Mean location and replicate dispersion answer different questions and must both be reported.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Continue through verified routes
Begin with the Polymer Science root, Gold Standard of Education and How X Works. Continue through Tensorial Entanglement Recovery, Eigenframes and Direction-Dependent Overshoot, Waiting-Time Spectra, Restart Superposition and Directional Recovery and Orthogonal Double-Step Strain, Tensor Memory and Cross Components.
Primary reading includes Edelman, Arias and Smith’s geometry of algorithms with orthogonality constraints, Davis and Kahan’s rotation of invariant subspaces, and Hameduddin and colleagues’ geometric decomposition of the conformation tensor. The replicate ledgers are original teaching examples.
Top · Contents · Previous · How X Works