Chapter map

1. 1. Base matrix
2. 2. Perturbation
3. 3. Norm
4. 4. Sum
5. 5. New singular values
6. 6. First change
7. 7. Second change
8. 8. Coordinatewise check
9. 9. Rectangular extension
10. 10. Dilation
11. 11. Aggregate norm
12. 12. Rank threshold
13. 13. Gap
14. 14. Application
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 step deserves a careful pause. Set A=diag(5,2).

A concrete situation

Its singular values are 5 and 2.

What the mathematics does not license

Order descending. A trustworthy workflow records the choice before seeing the final test result.

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. Perturbation

The useful question is not only how to calculate, but what the calculation assumes. Set E=diag(.3,-.1).

A concrete situation

The dimensions match.

What the mathematics does not license

Record signs. Evidence becomes stronger when the limitation is tested instead of hidden.

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.

---

3. Norm

A small example makes the mechanism visible. Compute ||E||2.

A concrete situation

It is .3.

What the mathematics does not license

Use maximum singular value. Keeping this boundary clear makes the result easier to audit and transfer.

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.

---

4. Sum

This is where modelling choices begin to affect evidence. Form A+E.

A concrete situation

It is diag(5.3,1.9).

What the mathematics does not license

Check arithmetic. A trustworthy workflow records the choice before seeing the final test result.

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.

---

5. New singular values

The central mathematical idea is concrete. Take absolute diagonal entries.

A concrete situation

They are 5.3 and 1.9.

What the mathematics does not license

Order descending. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record list. 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. First change

Here the geometry becomes operational. Subtract sigma1(A).

A concrete situation

The change is .3.

What the mathematics does not license

Take magnitude. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record change. 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. Second change

This step deserves a careful pause. Subtract sigma2(A).

A concrete situation

The change is -.1.

What the mathematics does not license

Magnitude is .1. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Record change. 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. Coordinatewise check

The useful question is not only how to calculate, but what the calculation assumes. Compare both with .3.

A concrete situation

Both satisfy the bound.

What the mathematics does not license

First reaches equality. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record result. 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. Rectangular extension

A small example makes the mechanism visible. Add a zero row and repeat.

A concrete situation

The nonzero singular values persist.

What the mathematics does not license

Do not require squareness. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record extension. 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. Dilation

This is where modelling choices begin to affect evidence. Build [[0,A],[A*,0]].

A concrete situation

Eigenvalues encode +/- singular values.

What the mathematics does not license

Track dimensions. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Record bridge. 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. Aggregate norm

The central mathematical idea is concrete. Compute a Frobenius error if needed.

A concrete situation

It gives a different summary.

What the mathematics does not license

Name the norm. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record distinction. 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. Rank threshold

Here the geometry becomes operational. Compare sigma_k with the noise scale.

A concrete situation

A margin supports robustness.

What the mathematics does not license

Do not make automatic rank claims. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record margin. 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. Gap

This step deserves a careful pause. Inspect adjacent singular values.

A concrete situation

Vectors need a gap theorem.

What the mathematics does not license

Do not infer direction stability. A trustworthy workflow records the choice before seeing the final test result.

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.

---

14. Application

The useful question is not only how to calculate, but what the calculation assumes. Bound E from measurement or approximation evidence.

A concrete situation

Then transfer to singular values.

What the mathematics does not license

Keep evidence separate. Evidence becomes stronger when the limitation is tested instead of hidden.

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

A small example makes the mechanism visible. State what missing data or model bias remains.

A concrete situation

The theorem is deterministic.

What the mathematics does not license

Do not imply fairness. Keeping this boundary clear makes the result easier to audit and transfer.

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 is where modelling choices begin to affect evidence. State dimensions, ordering and operator norm.

A concrete situation

The checked changes are .3 and .1.

What the mathematics does not license

Keep the conclusion precise. A trustworthy workflow records the choice before seeing the final test result.

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

Singular-value perturbation bounds show that every ordered singular value is Lipschitz continuous with respect to the operator norm. For matrices A and A+E of the same dimensions, |sigma_k(A+E)-sigma_k(A)| is at most ||E||2 for every valid k.

Unlike eigenvalues of a general matrix, singular values are nonnegative and exist for rectangular matrices. They measure stretch along orthogonal input and output directions, which makes them useful for data compression, inverse problems, least squares and low-rank approximation.

Let A=diag(5,2) and E=diag(0.3,-0.1). Then A+E=diag(5.3,1.9). The ordered singular-value shifts are 0.3 and -0.1, while ||E||2=0.3. The first singular value reaches the bound and the second remains inside it.

The same theorem applies when E is not diagonal and when A is rectangular. A diagonal example exposes the arithmetic, while the general proof relies on variational characterisations or Hermitian dilation. This is another reason that a simple check and a proof should be labelled separately.

Mirsky’s broader framework relates unitarily invariant norms to symmetric gauge functions of singular values. It allows vector-level comparisons between entire singular-value lists. The one-by-one operator-norm bound is especially easy to interpret, while Frobenius or Schatten norms can support aggregate error statements.

A small singular-value error does not automatically give a small singular-vector error. If two singular values are close, their singular subspaces may rotate significantly. Singular-vector stability needs a gap and a theorem such as Wedin’s, just as eigenvector stability needs separate spectral information.

In low-rank modelling, a bound can help decide whether an observed singular value is safely above the noise scale. It cannot prove that a chosen rank is scientifically meaningful, that missing data are random or that downstream predictions are fair. Those questions require additional evidence.

Students can practise by computing a residual matrix, its operator norm, the two singular-value lists and every coordinatewise difference. They should then compare an operator-norm statement with a Frobenius-norm statement and explain why the numerical conclusions differ.

Did You Know? Hermitian dilation places a rectangular matrix into a larger self-adjoint block matrix whose eigenvalues are the positive and negative singular values. This bridge lets many eigenvalue arguments become singular-value arguments without pretending the original matrix was square.

A second worked route uses Hermitian dilation. For a rectangular A, form the block matrix H(A)=[[0,A],[A*,0]]. Its positive eigenvalues are the singular values of A, its negative eigenvalues are their negatives, and extra zeros account for unequal dimensions. The perturbation H(E) has operator norm ||E||2. Applying a Hermitian eigenvalue perturbation statement to these blocks recovers coordinatewise singular-value stability. This derivation reveals why the bound is dimension compatible and why the operator norm is the natural right-hand side.

In an application, report more than the largest singular value. List the singular values near the proposed rank cutoff, the perturbation norm and the margin between signal candidates and the error scale. If the kth singular value lies only slightly above the bound, the theorem supports caution rather than a confident rank decision. If a stable subspace is needed, record the adjacent gap and use an eligible singular-subspace perturbation result. For students, this is a valuable model of responsible mathematics: one theorem controls values, another controls directions, and neither decides whether the data-generating assumptions or final interpretation are suitable.

The bound is also a calibration tool for computation. If an iterative algorithm returns A-hat with a certified residual or approximation error E=A-hat-A, the singular values of A-hat inherit an immediate uncertainty band of width ||E||2. That band is an error guarantee, not an instruction to round every value to the same number of decimals. Report the computed value, the norm ceiling and any independent numerical tolerance separately. This allows a reader to distinguish algorithmic approximation, measurement noise and floating-point effects instead of combining them into one vague statement.

---

Frequently asked questions

What does the operator-norm bound say?

Every ordered singular value changes by at most ||E||₂.

Can the matrices be rectangular?

Yes, provided A and A+E have the same dimensions.

Does it control singular vectors?

No. Vector or subspace stability needs a singular-value gap and another theorem.

What does Mirsky add?

A framework for comparing singular-value vectors through unitarily invariant norms.

Does a noise threshold select rank automatically?

No. Rank choice also depends on the model, loss and evidence.

What must be reported?

Dimensions, singular-value order, norm, perturbation source and whether the claim is coordinatewise or aggregate.

---

Primary source and fact-check note

L. Mirsky, “Symmetric Gauge Functions and Unitarily Invariant Norms,” The Quarterly Journal of Mathematics 11, 1960. 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

• Hermitian dilation
• Weyl eigenvalue source
• 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

Dimension audit

Keep A and A+E the same shape.

Order audit

Compare singular values in descending order.

Norm audit

Distinguish operator, Frobenius and Schatten norms.

Vector audit

Require a separate gap theorem for singular directions.

Mechanism workshop: calculate, challenge and transfer

Manual reconstruction

Build the smallest matrix example for rectangular singular-value stability. 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(5,2). The checked result is: Its singular values are 5 and 2. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Order descending. 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: Perturbation

Step 2 isolates perturbation. Set E=diag(.3,-.1). The checked result is: The dimensions match. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Record signs. 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 3: Norm

Step 3 isolates norm. Compute ||E||2. The checked result is: It is .3. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Use maximum singular value. 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 4: Sum

Step 4 isolates sum. Form A+E. The checked result is: It is diag(5.3,1.9). Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Check arithmetic. 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 5: New singular values

Step 5 isolates new singular values. Take absolute diagonal entries. The checked result is: They are 5.3 and 1.9. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Order descending. Record list. 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: First change

Step 6 isolates first change. Subtract sigma1(A). The checked result is: The change is .3. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Take magnitude. Record change. 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: Second change

Step 7 isolates second change. Subtract sigma2(A). The checked result is: The change is -.1. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Magnitude is .1. Record change. 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: Coordinatewise check

Step 8 isolates coordinatewise check. Compare both with .3. The checked result is: Both satisfy the bound. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: First reaches equality. Record result. 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: Rectangular extension

Step 9 isolates rectangular extension. Add a zero row and repeat. The checked result is: The nonzero singular values persist. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Do not require squareness. Record extension. 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: Dilation

Step 10 isolates dilation. Build [[0,A],[A*,0]]. The checked result is: Eigenvalues encode +/- singular values. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Track dimensions. Record bridge. 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: Aggregate norm

Step 11 isolates aggregate norm. Compute a Frobenius error if needed. The checked result is: It gives a different summary. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Name the norm. Record distinction. 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: Rank threshold

Step 12 isolates rank threshold. Compare sigma_k with the noise scale. The checked result is: A margin supports robustness. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Do not make automatic rank claims. Record margin. 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: Gap

Step 13 isolates gap. Inspect adjacent singular values. The checked result is: Vectors need a gap theorem. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Do not infer direction stability. 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 14: Application

Step 14 isolates application. Bound E from measurement or approximation evidence. The checked result is: Then transfer to singular values. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Keep evidence separate. 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. State what missing data or model bias remains. The checked result is: The theorem is deterministic. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Do not imply fairness. 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 dimensions, ordering and operator norm. The checked result is: The checked changes are .3 and .1. Preserve the matrix dimensions, ordering convention, norm and theorem hypotheses beside every number. The safeguard is: Keep the conclusion precise. 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

Eigenvalue substitution

Singular values are not general-matrix eigenvalues. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Wrong shape

The compared matrices must have compatible dimensions. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Unsorted lists

Use ordered singular values with multiplicities. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Norm ambiguity

Name operator or aggregate norm explicitly. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Vector promise

Stable values need not mean stable singular vectors. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Rank certainty

A noise bound does not choose a scientifically correct rank. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Data assumption

The theorem does not repair biased or missing data. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Prediction guarantee

Stable spectra do not guarantee downstream accuracy. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.