Chapter map

1. 1. Base matrix
2. 2. Gap
3. 3. Perturbation
4. 4. Norm
5. 5. Sum
6. 6. Angle equation
7. 7. Angle
8. 8. Sine
9. 9. Ratio
10. 10. Scale check
11. 11. Sign invariance
12. 12. Cluster
13. 13. Separation
14. 14. Data evidence
15. 15. Limit
16. 16. Conclusion

Visit the How Mathematics Works hub for the wider series and its connections to education, reasoning, science, computing, design and careers.

---

1. Base matrix

This is where modelling choices begin to affect evidence. Set A=diag(4,1).

A concrete situation

A is Hermitian.

What the mathematics does not license

Record spectrum. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Write A. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

2. Gap

The central mathematical idea is concrete. Subtract 1 from 4.

A concrete situation

The gap is 3.

What the mathematics does not license

Define the separated sets. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record gap. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

3. Perturbation

Here the geometry becomes operational. Set E=[[0,.2],[.2,0]].

A concrete situation

E is Hermitian.

What the mathematics does not license

Keep symmetry. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Write E. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

4. Norm

This step deserves a careful pause. Compute ||E||2.

A concrete situation

Its eigenvalues are +/-0.2, so norm is 0.2.

What the mathematics does not license

Use spectral norm. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record norm. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

5. Sum

The useful question is not only how to calculate, but what the calculation assumes. Form A+E.

A concrete situation

It is [[4,.2],[.2,1]].

What the mathematics does not license

Check entries. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record sum. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

6. Angle equation

A small example makes the mechanism visible. Use tan(2theta)=2b/(a-d).

A concrete situation

It equals 0.4/3.

What the mathematics does not license

Choose the top-vector branch. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Record equation. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

7. Angle

This is where modelling choices begin to affect evidence. Take half the arctangent.

A concrete situation

theta is about 0.06627 radians.

What the mathematics does not license

Do not confuse degrees. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record theta. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

8. Sine

The central mathematical idea is concrete. Compute sin theta.

A concrete situation

It is about 0.06622.

What the mathematics does not license

Round last. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record value. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

9. Ratio

Here the geometry becomes operational. Compute ||E||/gap.

A concrete situation

0.2/3 is about 0.06667.

What the mathematics does not license

Match theorem version. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Record ratio. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

10. Scale check

This step deserves a careful pause. Compare sine and ratio.

A concrete situation

The example has the predicted scale.

What the mathematics does not license

This is one case. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record check. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

11. Sign invariance

The useful question is not only how to calculate, but what the calculation assumes. Replace a vector by its negative.

A concrete situation

The subspace angle is unchanged.

What the mathematics does not license

Compare spans. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record invariance. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

12. Cluster

A small example makes the mechanism visible. Group repeated eigenvalues.

A concrete situation

Use principal angles for subspaces.

What the mathematics does not license

Do not pair arbitrary bases. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Record target. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

13. Separation

This is where modelling choices begin to affect evidence. Draw target and complement spectra.

A concrete situation

Use the eligible distance.

What the mathematics does not license

Do not choose a convenient gap. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record definition. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

14. Data evidence

The central mathematical idea is concrete. Bound the matrix error separately.

A concrete situation

A concentration theorem may supply it.

What the mathematics does not license

State probability level. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record source. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

15. Limit

Here the geometry becomes operational. Separate rotation from model selection.

A concrete situation

A stable subspace can still be irrelevant.

What the mathematics does not license

Do not overclaim. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Record limit. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

16. Conclusion

This step deserves a careful pause. State matrices, target subspace, gap and angle norm.

A concrete situation

The checked sine is about .06622.

What the mathematics does not license

Quote constants accurately. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Write conclusion. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

A seven-day practice lab

1. Day 1: define the decision, observation unit, X variables and Y or reference condition.
2. Day 2: reproduce the smallest worked example and check every dimension, unit and sign.
3. Day 3: build a transparent baseline and write what it assumes.
4. Day 4: fit the main method inside a training-only pipeline; record every parameter.
5. Day 5: challenge one assumption with a deliberately difficult dataset or operating regime.
6. Day 6: compare validation metrics with the practical workload created by errors.
7. Day 7: present one supported conclusion, one unresolved uncertainty and one safe next experiment.

Questions for a student, parent or mentor

• Which part of the model is a mathematical definition, and which part is a choice?
• What small calculation would catch a sign, scale or dimension error?
• What evidence comes from training, validation and the untouched test?
• What benign condition could look like a fault or prediction failure?
• How would the answer change for another response, asset, school project or time period?
• What should a human review before acting on the score?

A method-specific audit before deployment

The Davis–Kahan sin theta theorem studies how invariant subspaces of Hermitian matrices rotate under perturbation. Its central ratio compares the size of a residual or perturbation with a spectral separation. Small noise helps, but a clear eigengap is equally important.

For a single eigenvector, the relevant angle is taken between one-dimensional subspaces, so changing a vector’s sign does not change the answer. For several eigenvalues, principal angles compare whole subspaces. This language handles repeated eigenvalues more honestly than attempting to pair individual basis vectors.

Consider A=diag(4,1) and E=[[0,0.2],[0.2,0]]. The unperturbed top eigenvector is e1 and the original gap is 3. For A+E, the rotation angle theta of the top eigenvector satisfies tan(2 theta)=0.4/3. This gives theta about 0.06627 radians and sin theta about 0.06622.

The spectral norm of E is 0.2, and 0.2/3 is about 0.06667. The example therefore matches the scale predicted by a standard separated-eigenspace sin-theta bound. Constants and the precise gap definition depend on the theorem version, so an application must quote the version it actually uses.

If the gap shrinks while the perturbation stays fixed, the same matrix error can cause a much larger rotation. At an exact multiplicity, there may be no unique individual eigenvector to recover. The stable object is then the invariant subspace for the whole cluster.

This distinction matters in principal component analysis, spectral clustering, signal processing and numerical linear algebra. A stable leading subspace can be meaningful even when the basis vectors inside it rotate. A claim about one component needs more separation than a claim about their combined span.

The theorem does not certify that a population model is correct, that a sample covariance is unbiased or that a selected number of components is appropriate. It is a conditional perturbation result. The matrix estimation error and spectral gap must be justified by separate data or probability arguments.

Students can develop transfer by drawing both spectra on a number line, marking the target cluster, measuring its separation from the complement and then calculating the operator norm of the perturbation. This order prevents the common mistake of inserting whichever gap looks most convenient.

Did You Know? Subspace angles are encoded by singular values of products involving orthogonal projectors. That connection lets a geometric question about rotation be checked with the same singular-value tools used in matrix approximation.

A full classroom audit can compare three distinct quantities: the displacement of eigenvalues, the rotation of an eigenspace and the residual of an approximate eigenvector. Weyl-type bounds answer the first question; Davis–Kahan answers the second; residual versions of sin-theta results can address the third. Keeping these columns separate prevents a small eigenvalue error from being reported as a small directional error. Students should also test a nearly closed gap, where the same perturbation produces a visibly larger angle. The numerical experiment does not weaken the theorem; it illustrates why the denominator belongs in the statement and why spectral separation is evidence rather than decoration.

For a parent or younger student, the core habit is simple: reliability depends on both the size of the disturbance and the space between the choices. Two clearly separated directions are easier to distinguish than two nearly tied ones. That analogy is helpful only after the mathematical objects are named precisely. In a formal solution, draw the target cluster, its complement and the measured gap; report the matrix norm; specify whether the angle is measured by sine, tangent or a projector norm; and state the probability level if the matrix error came from random data. This sequence produces a claim that another reader can reproduce and challenge.

Another robust check compares projectors. If P and P-hat are orthogonal projectors onto equal-dimensional target subspaces, an operator norm of their difference captures the sine of the largest principal angle under standard conventions. This avoids sign choices and arbitrary rotations of a basis inside the same subspace. A student can verify that replacing an orthonormal basis Q by QR for an orthogonal matrix R leaves QQ* unchanged. That calculation explains why the theorem protects a geometric span rather than one preferred coordinate description of it.

---

Frequently asked questions

What does Davis–Kahan control?

It bounds angles between invariant subspaces of Hermitian matrices under separation conditions.

Why is an eigengap needed?

Without separation, nearby eigenvectors or subspaces can rotate substantially.

What happens with repeated eigenvalues?

The invariant subspace for the cluster is the meaningful object.

Is the constant always one?

The exact constant and denominator depend on the theorem version and separation setup.

Does stability validate a model?

No. It is conditional on the matrix error and spectral structure.

What must be reported?

Target subspaces, angle metric, perturbation norm, separation definition and theorem version.

---

Primary source and fact-check note

Chandler Davis and W. M. Kahan, “The Rotation of Eigenvectors by a Perturbation. III,” SIAM Journal on Numerical Analysis 7, 1970. Definitions and core claims in this article were checked against this primary publication. Calculations shown here were independently recomputed. Examples are educational illustrations; they are not reported industrial performance results.

Useful next reading

• Weyl eigenvalue source
• Minimax Laplace transform method
• Matrix Bernstein inequality
• How Mathematics Works
• Study and learning methods

Keep the options open. Mathematics education supports clearer questions, transferable problem-solving skills and informed choices across science, engineering, computing, AI, statistics, design and many other careers. It does not guarantee intelligence, income, admission or employment; those outcomes depend on far more than one subject or model.

Spectral audit: assumptions, norms and comparison logic

Target audit

Identify the invariant subspace or cluster.

Gap audit

Measure separation from the complementary spectrum.

Angle audit

Name the principal-angle norm being bounded.

Evidence audit

Justify the matrix error independently.

Mechanism workshop: calculate, challenge and transfer

Manual reconstruction

Build the smallest matrix example for eigenspace rotation. Label dimensions, spectra, norms and assumptions.

Assumption map

Place self-adjointness, diagonalizability, separation and norm conditions beside the step they permit.

Counterexample

Remove one hypothesis and identify the first invalid conclusion.

Sensitivity

Change one entry, angle, gap or condition number at a time.

Independent verification

Recompute spectra and norms through a second method.

Simulation

Sample several bounded perturbations while keeping the theorem distinct from observations.

Engineering transfer

Map the mechanism to covariance estimation, numerical solvers, compression or signal analysis.

AI transfer

Explain a spectral-learning use without promising model quality or employment.

Parent explanation

Describe the reasoning habit in plain language without guaranteed results.

Limit statement

Write one supported conclusion, one overclaim and one repaired version.

Evidence notebook: reconstruct every spectral step

Evidence notebook 1: Base matrix

Step 1 isolates base matrix. Set A=diag(4,1). The checked result is: A is Hermitian. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Record spectrum. Write A. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 2: Gap

Step 2 isolates gap. Subtract 1 from 4. The checked result is: The gap is 3. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Define the separated sets. Record gap. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 3: Perturbation

Step 3 isolates perturbation. Set E=[[0,.2],[.2,0]]. The checked result is: E is Hermitian. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Keep symmetry. Write E. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 4: Norm

Step 4 isolates norm. Compute ||E||2. The checked result is: Its eigenvalues are +/-0.2, so norm is 0.2. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Use spectral norm. Record norm. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 5: Sum

Step 5 isolates sum. Form A+E. The checked result is: It is [[4,.2],[.2,1]]. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Check entries. Record sum. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 6: Angle equation

Step 6 isolates angle equation. Use tan(2theta)=2b/(a-d). The checked result is: It equals 0.4/3. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Choose the top-vector branch. Record equation. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 7: Angle

Step 7 isolates angle. Take half the arctangent. The checked result is: theta is about 0.06627 radians. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Do not confuse degrees. Record theta. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 8: Sine

Step 8 isolates sine. Compute sin theta. The checked result is: It is about 0.06622. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Round last. Record value. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 9: Ratio

Step 9 isolates ratio. Compute ||E||/gap. The checked result is: 0.2/3 is about 0.06667. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Match theorem version. Record ratio. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 10: Scale check

Step 10 isolates scale check. Compare sine and ratio. The checked result is: The example has the predicted scale. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: This is one case. Record check. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 11: Sign invariance

Step 11 isolates sign invariance. Replace a vector by its negative. The checked result is: The subspace angle is unchanged. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Compare spans. Record invariance. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 12: Cluster

Step 12 isolates cluster. Group repeated eigenvalues. The checked result is: Use principal angles for subspaces. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Do not pair arbitrary bases. Record target. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 13: Separation

Step 13 isolates separation. Draw target and complement spectra. The checked result is: Use the eligible distance. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Do not choose a convenient gap. Record definition. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 14: Data evidence

Step 14 isolates data evidence. Bound the matrix error separately. The checked result is: A concentration theorem may supply it. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: State probability level. Record source. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 15: Limit

Step 15 isolates limit. Separate rotation from model selection. The checked result is: A stable subspace can still be irrelevant. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Do not overclaim. Record limit. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 16: Conclusion

Step 16 isolates conclusion. State matrices, target subspace, gap and angle norm. The checked result is: The checked sine is about .06622. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Quote constants accurately. Write conclusion. Change one input while holding the others fixed, recalculate the full spectrum or singular spectrum, and explain whether the change comes from size, gap, conditioning, basis geometry or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Failure cases and responsible recovery

Wrong gap

Use the theorem’s separation from the complementary spectrum. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Vector sign error

Compare subspaces, not signed representatives. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Cluster splitting

Repeated eigenvalues should be treated as a joint subspace. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Constant borrowing

Do not mix theorem versions. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Norm mismatch

Use the norm named by the result. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Error omission

The perturbation size still needs evidence. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Causation claim

Stable components do not prove a causal mechanism. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Model certainty

Subspace stability does not validate chosen features. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.