Choose your reading route
eduKateSG · Why Mathematics? · Factor-Graph Smoothing for Reliability Sensor Histories: Variables, Factors, Sparse Least Squares, Bayes Trees and Incremental Updates
Start with the question, audit the worked example, or jump to the assumptions and practice. Every chapter stays open for uninterrupted reading.
Understand the question
Define the quantity and follow the mechanism.Route 2 · Chapters 5–8
Reproduce the example
Audit the numbers one step at a time.Route 3 · Chapters 9–12
Check assumptions
See where design and calibration matter.Route 4 · Chapters 13–16
Connect to condition monitoring
Connect the method to a condition-monitoring investigation.Route 5 · Chapters 17–20
Study and verify
Practise, inspect failure cases and read further.
Full chapter index · Practice and checking · Mathematics Learning Hub
Why mathematics is important becomes tangible when a long sensor history contains thousands of states but each measurement touches only a few of them. A factor graph separates unknown variables from local probabilistic constraints. For Gaussian or linearised models, finding the most likely trajectory becomes sparse least squares. Bayes trees and incremental algorithms such as iSAM2 update only affected parts of the factorisation when new data arrive, rather than solving the entire history from scratch.
Consider two scalar states x₀ and x₁. Use a prior factor x₀≈0, a motion factor x₁−x₀≈1 and a measurement factor x₁≈2, all with unit variance. The objective is x₀²+(x₁−x₀−1)²+(2−x₁)². Setting derivatives to zero gives 2x₀−x₁+1=0 and 2x₁−x₀−3=0, so x₀=1/3 and x₁=5/3. Each residual is 1/3 in magnitude and total squared cost is 1/3. The graph makes each evidence source and its locality visible.
Full chapter index
Chapters 1–4 · Understand the question
Chapters 5–8 · Reproduce the example
Chapters 9–12 · Check assumptions
Chapters 13–16 · Connect to condition monitoring
Chapters 17–20 · Study and verify
Draw variables and local factors before forming a global matrix. Unknown states, parameters or landmarks become variable nodes. x zero and x one are two health states. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. A node is not an observation. Label time and units.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make variable nodes the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Did You Know? Sparse elimination order can change computing time dramatically without changing the underlying probability model or optimum. A small observation like this often reveals the conceptual boundary that a formula alone hides.
Each squared residual is one explicit probabilistic claim. Each prior, transition or measurement contributes a local probability or residual factor. Three factors define the teaching history. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. A missing factor means missing evidence, not zero evidence. Map every source.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make factor nodes the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Sparse elimination turns graph locality into computational savings. Negative log Gaussian factors become weighted squared residuals. The teaching variances are one. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. Non-Gaussian factors produce different costs. State the distribution.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make from probability to cost the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Incremental inference edits affected parts while preserving the factor record. The prior factor contributes x zero squared. It anchors the otherwise drifting chain. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. An overly tight prior can dominate data. Calibrate its variance.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make prior residual the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Primary research source: https://doi.org/10.1177/0278364911430419. Read its problem definition, mathematical criterion and assumptions before transferring the method to a reliability setting.
Draw variables and local factors before forming a global matrix. The transition factor contributes x one minus x zero minus one squared. It encodes expected degradation increment one. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. A deterministic-looking residual still has model uncertainty. Keep its weight visible.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make motion residual the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Did You Know? Sparse elimination order can change computing time dramatically without changing the underlying probability model or optimum. A small observation like this often reveals the conceptual boundary that a formula alone hides.
Each squared residual is one explicit probabilistic claim. The sensor factor contributes two minus x one squared. It pulls x one toward two. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. A biased measurement shifts both states through the chain. Test sensor bias.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make measurement residual the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Sparse elimination turns graph locality into computational savings. Differentiate the total cost and set both derivatives to zero. Equations are 2x zero minus x one plus one equals zero and 2x one minus x zero minus three equals zero. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. Nonlinear factors require iterative linearisation. Check gradients.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make normal equations the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Incremental inference edits affected parts while preserving the factor record. Solve the two simultaneous equations. x zero is one third and x one is five thirds. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. The solution compromises rather than satisfying every factor exactly. Inspect residuals.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make worked solution the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
| Quantity | Value |
|---|---|
| Prior residual | x0 |
| Motion residual | x1 – x0 – 1 |
| Measurement residual | 2 – x1 |
| Normal equation 1 | 2×0 – x1 + 1 = 0 |
| Normal equation 2 | 2×1 – x0 – 3 = 0 |
| Solution x0 | 1/3 |
| Solution x1 | 5/3 |
| Total squared cost | 1/3 |
Draw variables and local factors before forming a global matrix. Each residual magnitude is one third, so three squared residuals sum to one third. No single factor is ignored. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. Cost depends on weights and units. Nondimensionalise.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make worked cost the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Did You Know? Sparse elimination order can change computing time dramatically without changing the underlying probability model or optimum. A small observation like this often reveals the conceptual boundary that a formula alone hides.
Each squared residual is one explicit probabilistic claim. Local factors create sparse Jacobian and information matrices. Each motion factor touches neighbouring times only. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. Elimination can create fill-in. Choose ordering carefully.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make sparsity the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Sparse elimination turns graph locality into computational savings. Eliminate variables to factor the joint density into conditional pieces. A small chain becomes a triangular system. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. Poor ordering increases intermediate cliques. Profile memory and time.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make variable elimination the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Incremental inference edits affected parts while preserving the factor record. A Bayes tree organises conditional cliques so affected regions can be updated locally. A new late measurement edits only part of the tree. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. Large loop closures can affect much more. Measure update scope.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make bayes tree the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Supporting authoritative source: https://dspace.mit.edu/entities/publication/0cea54ca-7a8b-47f7-b4e9-337ab9295bad. Compare the published scope with this teaching example, especially the information set, numerical representation and conditions under which the result is claimed.
Draw variables and local factors before forming a global matrix. Reuse the previous factorisation when adding factors or variables. Streaming sensor data extend the history. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. Incremental does not mean approximate by necessity, but linearisation choices matter. Report thresholds.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make incremental updates the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Did You Know? Sparse elimination order can change computing time dramatically without changing the underlying probability model or optimum. A small observation like this often reveals the conceptual boundary that a formula alone hides.
Each squared residual is one explicit probabilistic claim. Nonlinear factors are relinearised when variable changes exceed chosen criteria. A new calibration shifts a sensor model. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. Too little relinearisation creates bias; too much costs time. Validate thresholds.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make relinearisation the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Sparse elimination turns graph locality into computational savings. Robust losses reduce the influence of outliers in least-squares graphs. A bad sensor packet is downweighted. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. Robust losses can also suppress real faults. Evaluate detection impact.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make robust factors the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Incremental inference edits affected parts while preserving the factor record. Remove old variables while retaining a summary factor on the remaining graph. A fixed-lag smoother bounds memory. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. Marginalisation can densify the summary and freeze linearisation. Audit approximations.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make marginalisation the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Draw variables and local factors before forming a global matrix. Inspect residuals by factor type, solver convergence, fill-in, condition numbers and coverage. One sensor family shows persistent bias. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. Low total cost can hide structured error. Plot residual groups.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make diagnostics the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Did You Know? Sparse elimination order can change computing time dramatically without changing the underlying probability model or optimum. A small observation like this often reveals the conceptual boundary that a formula alone hides.
Each squared residual is one explicit probabilistic claim. A graph combines inspections, dynamics and sensors into an auditable history. An onset window is linked to specific factors. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. The optimum remains model-conditioned. Combine with physical evidence.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make reliability decisions the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Sparse elimination turns graph locality into computational savings. The example connects graphs, simultaneous equations, squares and compromise. A learner can verify one third and five thirds. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. Graph means a dependency structure, not just a plotted curve. Draw nodes and factors.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make student and parent guidance the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Useful next reading: moving horizon estimation, information-form Kalman filtering, Mathematics Learning Hub, eduKate Mathematics Learning System, approach to learning mathematics, Bukit Timah Mathematics curriculum overview, cross-subject learning approach. Follow the link that answers the next question in your own reasoning, rather than treating the list as a reading quota.
Incremental inference edits affected parts while preserving the factor record. Factor graphs reveal which local statements build a global estimate and why sparse structure matters computationally. Change the measurement from two to three and resolve. The mathematics is useful only when the state, units, timing and information boundary stay attached to every number. Incremental smoothing still needs careful models and validation. State the changed factors.
A careful audit has four moves. First write the claim in ordinary language. Next preserve the intermediate quantities instead of jumping to the final estimate. Then check dimensions, signs, matrix validity or probability totals. Finally compare the result with the physical story and an alternative model. Save variable ordering, factor definitions and weights, residuals, linearisation points, elimination order, Bayes-tree changes, solver tolerances and marginalisation choices. This record lets another reader locate whether disagreement comes from data, assumptions, computation or interpretation.
For practice, make practice, misconceptions and next reading the focus and minimise x zero squared plus motion residual x one minus x zero minus one squared plus measurement residual two minus x one squared, solve the two normal equations to obtain x zero one third and x one five thirds with total cost one third. Predict the direction before using a calculator. Change one defensible input, recompute the result and explain why it moved. Finish with one supported conclusion, one uncertainty and one next check. That routine develops mathematical problem-solving skills that transfer beyond this one algorithm.
Quick answers
Why is mathematics important here? It turns a verbal concern about uncertainty, later evidence, shared information or robustness into quantities that can be checked and challenged.
Does a successful calculation prove the physical model? No. It shows what follows from the declared model, inputs and criterion. Validation against independent evidence remains necessary.
Which school mathematics transfers? Ratios, inequalities, functions, weighted averages, variance, graphs and matrices all reappear, joined by careful language about time and assumptions.
How should a student practise? Reproduce the small example, keep intermediate values, predict the direction of change and explain one limitation without jargon.
What should a practitioner report? Equations, data boundaries, initialization, uncertainty settings, numerical method, diagnostics, sensitivity tests and the rule that connects estimates to decisions.
How can a parent help? Ask what each number represents, when the information became available and what assumption would most change the conclusion. Explanation is part of mathematical mastery.
Extended practice laboratory
Practice case 1: Variable nodes
Setup: Use this chapter’s illustrative case: x zero and x one are two health states.
Task: Label time and units.
Checking target: A node is not an observation.
Practice case 2: Factor nodes
Setup: Use this chapter’s illustrative case: Three factors define the teaching history.
Task: Map every source.
Checking target: A missing factor means missing evidence, not zero evidence.
Practice case 3: From probability to cost
Setup: Use this chapter’s illustrative case: The teaching variances are one.
Task: State the distribution.
Checking target: Non-Gaussian factors produce different costs.
Practice case 4: Prior residual
Setup: Use this chapter’s illustrative case: It anchors the otherwise drifting chain.
Task: Calibrate its variance.
Checking target: An overly tight prior can dominate data.
Practice case 5: Motion residual
Setup: Use this chapter’s illustrative case: It encodes expected degradation increment one.
Task: Keep its weight visible.
Checking target: A deterministic-looking residual still has model uncertainty.
Practice case 6: Measurement residual
Setup: Use this chapter’s illustrative case: It pulls x one toward two.
Task: Test sensor bias.
Checking target: A biased measurement shifts both states through the chain.
Practice case 7: Normal equations
Setup: Use this chapter’s illustrative case: Equations are 2x zero minus x one plus one equals zero and 2x one minus x zero minus three equals zero.
Task: Check gradients.
Checking target: Nonlinear factors require iterative linearisation.
Practice case 8: Worked solution
Setup: Use this chapter’s illustrative case: x zero is one third and x one is five thirds.
Task: Inspect residuals.
Checking target: The solution compromises rather than satisfying every factor exactly.
Practice case 9: Worked cost
Setup: Use this chapter’s illustrative case: No single factor is ignored.
Task: Nondimensionalise.
Checking target: Cost depends on weights and units.
Practice case 10: Sparsity
Setup: Use this chapter’s illustrative case: Each motion factor touches neighbouring times only.
Task: Choose ordering carefully.
Checking target: Elimination can create fill-in.