Choose your reading route
eduKateSG · Why Mathematics? · Normalized Innovation Squared for Kalman Filter Reliability: Mahalanobis Distance, Chi-Square Gates, Consistency and Fault Detection
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 6–10
Reproduce the example
Audit the numbers one step at a time.Route 3 · Chapters 11–15
Check assumptions
See where design and calibration matter.Route 4 · Chapters 16–20
Explore applications
Connect the method to a concrete investigation.Route 5 · Chapters 21–24
Study and verify
Practise, inspect failure cases and read further.
Full chapter index · Practice and checking · Mathematics Learning Hub
Why mathematics is important becomes clear when a residual must be judged relative to its predicted uncertainty. The normalized innovation squared (NIS) is the innovation’s squared Mahalanobis distance: ε=νᵀS⁻¹ν. Under a correctly specified linear Gaussian filter, it has a chi-square distribution with degrees of freedom equal to the measurement dimension. This turns residual checking into a calibrated consistency question rather than a raw-size guess.
For a scalar innovation ν=3 and innovation variance S=4, NIS=3²/4=2.25. The NIST 95% upper critical value for one chi-square degree of freedom is 3.841, so this single value is below that upper threshold. That does not prove the model is correct. Repeated values that are systematically too high or too low, serial correlation, bias and selection effects can still reveal inconsistency.
Did You Know? NIS can be computed online without knowing the true hidden state. That makes it useful for operational monitoring, but it tests measurement-domain consistency rather than every aspect of state accuracy.
This article answers one intent: The task is to decide whether innovations and their predicted covariance are mutually consistent, while controlling false alarms and avoiding threshold worship.
NIS is best understood as a conversation between a residual and its covariance. A residual says how far observation and prediction differ. The covariance says how much difference the model expected. Their quadratic combination produces a dimensionless distance that can be compared with a reference distribution. This is more informative than a fixed raw threshold because sensors with different scales and correlated errors can be assessed in a common probabilistic language.
The chi-square reference is conditional, not ceremonial. It depends on the measurement dimension and on assumptions about the innovation distribution and covariance calibration. Reusing one threshold after a sensor drops out, after measurements are gated or after an adaptive algorithm has tuned covariance from the same residuals can invalidate the interpretation. The article therefore treats dimensions, selection and dependence as part of the calculation. A table lookup is the final step of a chain, not the beginning.
For students, the topic links variance, vectors, matrices, probability and hypothesis tests. For a reliability team, it links live monitoring to an auditable false-alarm policy. Both groups should practise cautious conclusions. A value below 3.841 in the scalar 95% example means only that this one observation does not exceed that chosen upper boundary. It does not prove the sensor, state model or covariance is correct. Repeated behaviour, direction, timing and physical evidence complete the judgement.
Full chapter index
Chapters 1–5 · Understand the question
Chapters 6–10 · Reproduce the example
Chapters 11–15 · Check assumptions
Chapters 16–20 · Explore applications
Chapters 21–24 · Study and verify
Start with the mathematical object, because the name of a method is less useful than the quantity it changes. The same numerical residual can be ordinary under high uncertainty and alarming under low uncertainty. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
A three-unit error is assessed against S=4 rather than judged alone. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
A large S can hide poor model calibration. Inspect both residual and covariance. For transfer, change one input at a time, predict the direction first and then calculate. This reveals whether the idea is understood rather than memorised.
The practical value appears when the equation is connected to one observable decision. Innovation is observed measurement minus its predicted measurement before the update. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
ν=z−H x̂⁻ in a linear model. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Post-fit residuals have different covariance and interpretation. Label timing and equation. For responsible use, keep the raw observation beside the transformed value. A robust or normalised statistic should never erase the evidence from which it was formed.
This chapter separates the mechanism from the interpretation so that a calculation does not become an unsupported claim. S=H P⁻ Hᵀ+R combines predicted state uncertainty and measurement noise. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
Scalar S=4 sets the scale for the worked residual. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Missing cross-covariance or wrong R changes every NIS value. Audit the full covariance model. For checking, compare the result with a simpler baseline and one plausible alternative explanation. Agreement is informative; disagreement is a prompt to investigate.
A careful model keeps units, timing and assumptions visible before it produces a convenient number. NIS uses νᵀS⁻¹ν, accounting for scale and correlation. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
For scalar ν=3,S=4, ε=9/4=2.25. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Explicit inversion is unnecessary and can be unstable. Solve Sx=ν and compute νᵀx. For learning, explain the mechanism aloud in plain language, then point to the exact term in the equation that carries that meaning.
The next step is reproducibility: another reader should be able to rebuild the example from the stated inputs. Under stated linear Gaussian consistency conditions, NIS follows chi square with measurement dimension degrees of freedom. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
A scalar measurement uses one degree of freedom. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Nonlinearity, adaptation and gating can change the reference behaviour. Validate the distribution empirically. For reporting, separate what was observed, what the model assumed, what was calculated and what action remains a human or engineering decision.
Start with the mathematical object, because the name of a method is less useful than the quantity it changes. Compare ε=2.25 with the NIST 95% upper value 3.841 for one degree of freedom. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
The observation stays below that one-sided boundary. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Below threshold does not prove correctness. Treat the result as one piece of evidence. For transfer, change one input at a time, predict the direction first and then calculate. This reveals whether the idea is understood rather than memorised.
| Quantity | Value |
|---|---|
| Innovation ν | 3 |
| Innovation variance S | 4 |
| NIS ν²/S | 2.25 |
| χ² 95% upper value, 1 df | 3.841 |
| Single-point result | Below this upper threshold |
The practical value appears when the equation is connected to one observable decision. Values that are persistently too small can indicate inflated covariance, just as large values can indicate underestimated uncertainty. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
A filter reports huge S and produces NIS near zero. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Only using an upper gate misses over-conservatism. Use lower and upper checks when appropriate. For responsible use, keep the raw observation beside the transformed value. A robust or normalised statistic should never erase the evidence from which it was formed.
This chapter separates the mechanism from the interpretation so that a calculation does not become an unsupported claim. A 5% upper-tail rule expects some exceedances even when assumptions hold. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
About one in twenty independent scalar tests exceeds the 95% boundary in the long run. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Sequential dependence and repeated testing complicate this expectation. Plan false-alarm handling. For checking, compare the result with a simpler baseline and one plausible alternative explanation. Agreement is informative; disagreement is a prompt to investigate.
A careful model keeps units, timing and assumptions visible before it produces a convenient number. With m measurements, degrees of freedom are m when the full vector and covariance are valid. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
A three-channel innovation uses a chi-square reference with three degrees of freedom. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Missing channels change m and the threshold. Compute the active dimension each epoch. For learning, explain the mechanism aloud in plain language, then point to the exact term in the equation that carries that meaning.
The next step is reproducibility: another reader should be able to rebuild the example from the stated inputs. Summing independent NIS values over N epochs yields a chi-square reference with Nm degrees of freedom under assumptions. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
A window provides more stable evidence than one point. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Overlapping windows and autocorrelation reduce independence. Use correct degrees of freedom or simulation. For reporting, separate what was observed, what the model assumed, what was calculated and what action remains a human or engineering decision.
Start with the mathematical object, because the name of a method is less useful than the quantity it changes. NIS is available without a truth trajectory, making it practical for live filter monitoring. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
A dashboard tracks NIS and its control limits. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
It cannot reveal all state-domain errors. Use truth-based NEES when trustworthy truth is available. For transfer, change one input at a time, predict the direction first and then calculate. This reveals whether the idea is understood rather than memorised.
The practical value appears when the equation is connected to one observable decision. Systematically high NIS can suggest underestimated Q or R, model bias or unmodelled dynamics. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
NASA work uses chi-square statistics to examine covariance realism and tuning. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Changing Q or R until points fit a band can overfit. Combine statistics with physical modelling. For responsible use, keep the raw observation beside the transformed value. A robust or normalised statistic should never erase the evidence from which it was formed.
This chapter separates the mechanism from the interpretation so that a calculation does not become an unsupported claim. A sudden NIS exceedance can trigger an inspection or alternate sensor logic. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
One corrupted packet produces a high value. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
The statistic does not localise the cause by itself. Use residual components and redundancy. For checking, compare the result with a simpler baseline and one plausible alternative explanation. Agreement is informative; disagreement is a prompt to investigate.
A careful model keeps units, timing and assumptions visible before it produces a convenient number. A validation gate can reject or downweight measurements with excessive Mahalanobis distance. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
An association system excludes observations outside an ellipsoid. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Gating changes the observed residual distribution through selection. Account for truncation in later diagnostics. For learning, explain the mechanism aloud in plain language, then point to the exact term in the equation that carries that meaning.
The next step is reproducibility: another reader should be able to rebuild the example from the stated inputs. A consistent white innovation sequence should not show systematic autocorrelation. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
NIS can look acceptable while residual signs alternate predictably. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Magnitude testing alone misses temporal structure. Plot autocorrelation and run whiteness checks. For reporting, separate what was observed, what the model assumed, what was calculated and what action remains a human or engineering decision.
Start with the mathematical object, because the name of a method is less useful than the quantity it changes. NIS squares the innovation, so it does not show direction. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
Repeated positive innovations can signal bias even if each NIS is modest. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
A two-sided magnitude test may miss persistent sign structure. Monitor means and cumulative sums. For transfer, change one input at a time, predict the direction first and then calculate. This reveals whether the idea is understood rather than memorised.
The practical value appears when the equation is connected to one observable decision. EKF, UKF and particle approximations can use NIS-like diagnostics, but the ideal chi-square reference may be approximate. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
Linearisation error broadens or distorts the empirical distribution. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Applying textbook thresholds blindly creates false confidence. Calibrate through simulation and field residuals. For responsible use, keep the raw observation beside the transformed value. A robust or normalised statistic should never erase the evidence from which it was formed.
This chapter separates the mechanism from the interpretation so that a calculation does not become an unsupported claim. When Q and R are estimated from the same innovations, diagnostic and tuning data are no longer independent. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
An adaptive rule can force NIS toward target values. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
A pleasing NIS plot may reflect circular tuning. Use held-out periods and alternative diagnostics. For checking, compare the result with a simpler baseline and one plausible alternative explanation. Agreement is informative; disagreement is a prompt to investigate.
A careful model keeps units, timing and assumptions visible before it produces a convenient number. NIS can support sensor health monitoring and covariance-aware alarms. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
A high value prompts cross-checking rather than immediate component condemnation. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Safety decisions need consequence analysis and validated rules. Separate statistical trigger from authorised action. For learning, explain the mechanism aloud in plain language, then point to the exact term in the equation that carries that meaning.
The next step is reproducibility: another reader should be able to rebuild the example from the stated inputs. The calculation connects vectors, matrices, variance, ratios and hypothesis testing. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
A learner can verify 2.25 and look up 3.841. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Below a threshold is not the same as proof of no problem. Practise cautious language. For reporting, separate what was observed, what the model assumed, what was calculated and what action remains a human or engineering decision.
Start with the mathematical object, because the name of a method is less useful than the quantity it changes. Store innovation, S, active dimension, NIS, thresholds, gating decision and model version. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
Unit tests include scalar and correlated-vector cases. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Hard-coded thresholds can become wrong when dimensions change. Compute them from declared probabilities. For transfer, change one input at a time, predict the direction first and then calculate. This reveals whether the idea is understood rather than memorised.
The practical value appears when the equation is connected to one observable decision. NIS diagnoses or gates inconsistency; Student’s t, Huber and correntropy methods change influence. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
A system may use NIS to trigger a robust mode. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Using the same residual for selection and performance claims can bias evaluation. Separate monitoring, adaptation and validation datasets. For responsible use, keep the raw observation beside the transformed value. A robust or normalised statistic should never erase the evidence from which it was formed.
This chapter separates the mechanism from the interpretation so that a calculation does not become an unsupported claim. NIS is a model-consistency statistic, not a universal fault probability. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
Change S from 4 to 1 and 9 while holding ν=3. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
A threshold exceedance has several possible causes. List at least three before acting. For checking, compare the result with a simpler baseline and one plausible alternative explanation. Agreement is informative; disagreement is a prompt to investigate.
A careful model keeps units, timing and assumptions visible before it produces a convenient number. Primary research connects NIS with online consistency tests; NIST supplies chi-square critical values; NASA documents filter-tuning use. The importance of mathematics here lies in converting that statement into a quantity that can be checked, challenged and recomputed.
Neighbouring eduKate articles explain adaptive filtering, strong tracking and sensor fusion. Work through the numbers or logic in the order shown, retain the units and ask whether the direction of change makes sense. A result that cannot survive this basic audit should not be carried into a reliability claim.
Threshold tables remain conditional on assumptions. Verify dimensions, probability and dependence. For learning, explain the mechanism aloud in plain language, then point to the exact term in the equation that carries that meaning.
The nearest published companions are Schmidt–Kalman filtering, adaptive Kalman filtering, strong tracking filtering and factor-graph smoothing. Use the article whose mechanism matches the uncertainty you actually face. The eduKateSG Mathematics Learning Hub, eduKateSingapore Mathematics World, eduKate Sengkang Mathematics Hub, Bukit Timah Tutor Singapore Mathematics Hub and eduKate Punggol Mathematics Pathway provide wider routes by level, purpose and learner need. The primary sources used for this article are https://doi.org/10.1002/acs.2571, https://www.itl.nist.gov/div898/handbook/eda/section3/eda3674.htm, https://ntrs.nasa.gov/citations/20170003685.
Frequently asked questions
What is normalized innovation squared?
NIS is the squared Mahalanobis distance of an innovation: ε=νᵀS⁻¹ν. It judges residual size relative to predicted measurement uncertainty and correlation.
Why is NIS compared with chi-square values?
Under a correctly specified linear Gaussian filter, NIS follows a chi-square distribution with degrees of freedom equal to the active measurement dimension. This creates conditional consistency limits.
Does an NIS below the upper threshold prove the filter is correct?
No. It says only that the tested value does not exceed that chosen boundary. Bias, serial correlation, overlarge covariance and other model errors may remain.
Can NIS be used when the true state is unknown?
Yes. It uses measurement innovations and their predicted covariance, so it is available online without truth. It does not test every aspect of state-domain accuracy.
What changes when a sensor is missing?
The active measurement dimension and innovation covariance change, so the degrees of freedom and threshold must also change. Hard-coded limits can become invalid.
Extended practice laboratory
Practice case 1: Scalar NIS
Setup: Use ν=3,S=4.
Checking target: Confirm ε=2.25.
Practice case 2: Variance sensitivity
Setup: Use S=1,4,9 with ν=3.
Checking target: Explain the scale effect.
Practice case 3: Critical value
Setup: Look up one-degree 95% upper value.
Checking target: Confirm 3.841.
Practice case 4: Expected exceedances
Setup: Imagine 200 independent tests.
Checking target: Estimate about ten upper exceedances at 5%.
Practice case 5: Vector case
Setup: Use a two-dimensional ν and correlated S.
Checking target: Solve rather than invert.
Practice case 6: Dimension change
Setup: Drop one missing sensor.
Checking target: Update degrees of freedom and threshold.
Practice case 7: Window sum
Setup: Sum ten scalar NIS values.
Checking target: Use ten degrees of freedom under independence.
Practice case 8: Bias sequence
Setup: Create modest positive innovations.
Checking target: Compare mean monitoring with NIS.
Practice case 9: Autocorrelation
Setup: Alternate residual signs with constant magnitude.
Checking target: Show why NIS alone misses timing.
Practice case 10: Decision memo
Setup: Explain one threshold exceedance.
Checking target: List causes, evidence and next check.