Chapter map

1. 1. Matrix
2. 2. Target
3. 3. Probe
4. 4. Sample
5. 5. Sample norm
6. 6. Basis
7. 7. Projection
8. 8. Matrix energy
9. 9. Captured energy
10. 10. Residual square
11. 11. Residual norm
12. 12. Optimal error
13. 13. Comparison
14. 14. Oversampling
15. 15. Power iteration
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. Matrix

This step deserves a careful pause. Set A=diag(6,3,1).

A concrete situation

Singular values are 6,3,1.

What the mathematics does not license

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

Try it, check it, transfer it

Write matrix. 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. Target

The useful question is not only how to calculate, but what the calculation assumes. Choose rank one.

A concrete situation

The optimum retains 6.

What the mathematics does not license

State target. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record k. 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. Probe

A small example makes the mechanism visible. Set omega=(1,1,1).

A concrete situation

This is one illustrative draw.

What the mathematics does not license

Not a distributional proof. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record probe. 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. Sample

This is where modelling choices begin to affect evidence. Compute y=A omega.

A concrete situation

y=(6,3,1).

What the mathematics does not license

Multiply entries. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Record y. 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. Sample norm

The central mathematical idea is concrete. Compute ||y||2^2.

A concrete situation

The value is 46.

What the mathematics does not license

Use 36+9+1. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record 46. 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. Basis

Here the geometry becomes operational. Set q=y/sqrt46.

A concrete situation

q is unit length.

What the mathematics does not license

Verify normalization. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record q. 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. Projection

This step deserves a careful pause. Form qq* A.

A concrete situation

This is rank at most one.

What the mathematics does not license

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

Try it, check it, transfer it

Record approximation. 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. Matrix energy

The useful question is not only how to calculate, but what the calculation assumes. Compute ||A||F^2.

A concrete situation

The value is 46.

What the mathematics does not license

Use singular values. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record total. 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. Captured energy

A small example makes the mechanism visible. Compute ||q* A||2^2.

A concrete situation

1378/46≈29.9565.

What the mathematics does not license

Check 36^2+9^2+1. Keeping this boundary clear makes the result easier to audit and transfer.

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.

---

10. Residual square

This is where modelling choices begin to affect evidence. Subtract captured from total.

A concrete situation

About 16.0435.

What the mathematics does not license

Use orthogonal projection. A trustworthy workflow records the choice before seeing the final test result.

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.

---

11. Residual norm

The central mathematical idea is concrete. Take square root.

A concrete situation

About 4.0054.

What the mathematics does not license

Round last. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record error. 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. Optimal error

Here the geometry becomes operational. Discard 3 and 1 after exact SVD.

A concrete situation

sqrt10≈3.1623.

What the mathematics does not license

Use Eckart–Young. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record baseline. 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. Comparison

This step deserves a careful pause. Compare 4.0054 with 3.1623.

A concrete situation

One probe is worse but informative.

What the mathematics does not license

Do not claim optimality. 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. Oversampling

The useful question is not only how to calculate, but what the calculation assumes. Use more random columns.

A concrete situation

Failure risk drops.

What the mathematics does not license

State parameter. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record plan. 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. Power iteration

A small example makes the mechanism visible. Apply spectral amplification if needed.

A concrete situation

Extra passes cost time.

What the mathematics does not license

State q passes. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record tradeoff. 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 randomness, rank, passes and residual.

A concrete situation

The worked error is about 4.0054.

What the mathematics does not license

Keep probability claim 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

Randomized SVD algorithms use random test vectors to find a low-dimensional subspace that captures much of a matrix’s action. The large matrix is then compressed to that subspace, and a smaller deterministic factorization produces an approximate SVD.

Take A=diag(6,3,1) and one test vector omega=(1,1,1). The sample y=A omega=(6,3,1) has squared norm 46, so q=y/sqrt(46) is a unit vector biased toward the leading singular direction.

Projecting A onto span(q) gives q q* A. The squared Frobenius residual is ||A||_F^2-||q* A||_2^2. Here ||A||_F^2=46 and ||q* A||_2^2=(36^2+9^2+1^2)/46=1378/46≈29.9565.

The residual Frobenius error is therefore sqrt(46-1378/46)≈4.0054. The optimal rank-one truncated-SVD error is sqrt(3^2+1^2)=sqrt(10)≈3.1623. One unoversampled random probe captures useful structure but need not be optimal.

Practical algorithms use several test vectors and oversampling. Power iterations apply A A* or A* A repeatedly to amplify spectral separation, improving accuracy when singular values decay slowly. Each extra pass costs time and data access.

Randomization changes the guarantee. The algorithm returns an approximation whose error is controlled in expectation or with high probability under the stated sampling scheme. A seed and the number of passes belong in a reproducible report.

The method is valuable for matrices too large for a full deterministic SVD, including scientific computing, image analysis and large data systems. It can reduce communication and exploit modern hardware, but it does not make computation free.

A sketch must preserve the subspace relevant to the target. If the sampling distribution, dimensions or numerical precision are unsuitable, important directions can be missed. Oversampling reduces risk; it does not remove the need to verify residuals.

Students can compare three objects: the random range, the compressed matrix and the final approximate factors. Keeping these stages separate reveals where error enters and where a deterministic calculation resumes.

Did You Know? Randomized linear algebra often uses probability to reduce computation while still producing explicit residuals that can be checked after the run. Random choice proposes a subspace; mathematics audits the result.

A responsible application reports target rank, oversampling, power iterations, random seed policy, norm of the residual, comparison with a baseline and the probability statement from the theorem used. It should not claim that one successful run proves universal accuracy.

With a sketch matrix Omega containing l=k+p columns, the oversampling parameter p gives the sampled range extra room to capture dominant directions. After Y=A Omega is formed, a QR factorization produces Q, and the smaller matrix B=Q* A is factored. The final approximation Q B remains tied to the sampled range, so the residual ||A-QQ*A|| is a direct diagnostic of sketch quality.

Power iteration replaces Y=A Omega with Y=(A A*)^q A Omega, usually with re-orthogonalization for numerical stability. Each singular value is effectively raised to the power 2q+1 before the range is identified, widening relative separation. This can help a slow-decay spectrum, but very large powers without re-orthogonalization may lose smaller directions to floating-point effects.

The choice between spectral and Frobenius error matters. A spectral guarantee protects the worst missed direction, while a Frobenius guarantee combines all missed directions. Randomized methods often provide both kinds of bounds with different constants and tail statements. Quote the eligible theorem instead of moving one error formula into the other.

For streaming or pass-limited data, the number of times the algorithm reads A can dominate arithmetic cost. A method with a slightly looser approximation may be preferable if it saves expensive passes. That is an engineering tradeoff supported by mathematics, not a claim that randomization is always faster.

Students can repeat the diagonal example with probes (1,0,0), (0,1,0) and several Gaussian draws. The first probe finds the optimal leading direction, the second misses it completely, and random draws fall between. This contrast explains why probability, oversampling and residual verification work together.

A final reproducibility check estimates the residual without forming a dense residual matrix when A is enormous. Independent random test vectors can probe (I-QQ*)A, while deterministic applications may use available operator-norm estimators or exact smaller products. The estimator introduces its own uncertainty, so the report should distinguish the approximation algorithm’s random sketch from the residual-checking procedure. When the data are sensitive, publish parameters and aggregate diagnostics without exposing records. This keeps computational efficiency, evidence and privacy as separate design questions.

---

Frequently asked questions

What is randomized SVD?

A range-finding sketch followed by a smaller deterministic factorization.

Why use random probes?

They can capture a dominant subspace with fewer passes over a large matrix.

Is the result optimal?

Not necessarily; it is compared with the best truncated-SVD error.

What improve accuracy?

Oversampling and power iterations, at additional cost.

What must be reproducible?

Random scheme, seed policy, rank, oversampling, passes, norm and residual.

Does one run prove accuracy?

No. Use a theorem-level probability statement and check the realised residual.

---

Primary source and fact-check note

Nathan Halko, Per-Gunnar Martinsson and Joel A. Tropp, “Finding Structure with Randomness,” SIAM Review 53, 2011. 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

• Eckart–Young–Mirsky theorem
• Matrix Gaussian series
• Singular-value perturbation
• 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 optimization logic

Sketch audit

State random distribution and dimensions.

Rank audit

Record target rank and oversampling.

Pass audit

Count power iterations and data passes.

Residual audit

Verify the realised approximation independently.

Mechanism workshop: calculate, challenge and transfer

Manual reconstruction

Build the smallest matrix example for random range finding. Label spectra, shapes, norms and assumptions.

Assumption map

Place adjoints, p, rank, measurement and randomness conditions beside the step they permit.

Counterexample

Remove one hypothesis and identify the first invalid conclusion.

Sensitivity

Change one singular value, alignment, measurement or random probe at a time.

Independent verification

Recompute traces, spectra, ranks and residuals through a second route.

Simulation

Repeat a random sketch or noisy recovery while keeping observation and theorem separate.

Engineering transfer

Map the mechanism to compression, control, signal processing or large data.

AI transfer

Explain a learning use without promising accuracy, intelligence or employment.

Parent explanation

Describe the reasoning habit in plain language without guaranteed results.

Limit statement

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

Evidence notebook: reconstruct every matrix step

Evidence notebook 1: Matrix

Step 1 isolates matrix. Set A=diag(6,3,1). The checked result is: Singular values are 6,3,1. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: Record A. Write matrix. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 2: Target

Step 2 isolates target. Choose rank one. The checked result is: The optimum retains 6. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: State target. Record k. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 3: Probe

Step 3 isolates probe. Set omega=(1,1,1). The checked result is: This is one illustrative draw. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: Not a distributional proof. Record probe. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 4: Sample

Step 4 isolates sample. Compute y=A omega. The checked result is: y=(6,3,1). Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: Multiply entries. Record y. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 5: Sample norm

Step 5 isolates sample norm. Compute ||y||2^2. The checked result is: The value is 46. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: Use 36+9+1. Record 46. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 6: Basis

Step 6 isolates basis. Set q=y/sqrt46. The checked result is: q is unit length. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: Verify normalization. Record q. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 7: Projection

Step 7 isolates projection. Form qq* A. The checked result is: This is rank at most one. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: Track shapes. Record approximation. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 8: Matrix energy

Step 8 isolates matrix energy. Compute ||A||F^2. The checked result is: The value is 46. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: Use singular values. Record total. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 9: Captured energy

Step 9 isolates captured energy. Compute ||q* A||2^2. The checked result is: 1378/46≈29.9565. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: Check 36^2+9^2+1. Record value. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 10: Residual square

Step 10 isolates residual square. Subtract captured from total. The checked result is: About 16.0435. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: Use orthogonal projection. Record value. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 11: Residual norm

Step 11 isolates residual norm. Take square root. The checked result is: About 4.0054. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: Round last. Record error. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 12: Optimal error

Step 12 isolates optimal error. Discard 3 and 1 after exact SVD. The checked result is: sqrt10≈3.1623. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: Use Eckart–Young. Record baseline. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 13: Comparison

Step 13 isolates comparison. Compare 4.0054 with 3.1623. The checked result is: One probe is worse but informative. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: Do not claim optimality. Record gap. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 14: Oversampling

Step 14 isolates oversampling. Use more random columns. The checked result is: Failure risk drops. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: State parameter. Record plan. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension or rounding. Ask a second reader to reproduce the calculation. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 15: Power iteration

Step 15 isolates power iteration. Apply spectral amplification if needed. The checked result is: Extra passes cost time. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: State q passes. Record tradeoff. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension 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 randomness, rank, passes and residual. The checked result is: The worked error is about 4.0054. Preserve dimensions, singular values, p, rank, measurement map or random scheme beside every number. The safeguard is: Keep probability claim precise. Write conclusion. Change one input while holding the others fixed, recalculate traces, norms, ranks or residuals, and explain whether the change comes from magnitude, alignment, convex relaxation, sketch dimension 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

One-draw proof

An example is not a probability theorem. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

No oversampling

A minimal sketch can miss useful directions. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Pass omission

Power iteration costs data passes. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Optimality claim

Randomized approximation need not equal exact SVD. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Residual omission

A realised sketch should be checked. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Seed secrecy

Reproducibility needs a seed policy. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Slow-decay neglect

Flat spectra can require more work. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.

Speed guarantee

Hardware and access patterns affect runtime. Build a small counterexample, choose an eligible theorem and state the missing evidence before calculating again.