📊 statistics

Selection-Induced Contraction of Innovation Statistics in Gated Kalman Filters

This paper demonstrates that validation gating and nearest-neighbor association in Kalman filters fundamentally alter innovation statistics by inducing a deterministic, dimension-dependent contraction of the innovation covariance, proving that nominal statistical properties cannot be preserved under these selection processes.

Original authors: Barak Or

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Barak Or

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are running a high-tech radar system that tracks moving objects, like airplanes or satellites. To do this, the system uses a mathematical tool called a Kalman Filter. Think of this filter as a very smart guesser. It predicts where an object should be, looks at what the sensors actually see, and then calculates the difference between the two. This difference is called the innovation.

In a perfect world, these differences (innovations) would bounce around randomly, following a predictable pattern (like a bell curve). If the differences get too big, the system assumes the sensor made a mistake or there's a glitch, so it throws that data away. This process of throwing away "bad" or "too-big" data is called gating.

The paper by Barak Or reveals a surprising truth: The act of throwing away the "big" differences actually tricks the system into thinking the world is more perfect than it really is.

Here is the breakdown of the paper's findings using simple analogies:

1. The "Filter" That Filters Out the Truth

Imagine you are a teacher grading a class of students. You know that test scores usually follow a normal pattern: some get 50s, some 70s, some 90s.

The Paper's Claim: The teacher decides, "I'm only going to look at the papers where the score is between 80 and 100."
The Result: After looking only at those high scores, the teacher calculates the "average score" of the class. Because they threw away all the 50s and 60s, the new average is much higher than the real class average.
In the Radar: The radar throws away any measurement that is "too far off" from the prediction. The remaining measurements are all "close" to the prediction. If the system then looks at these remaining measurements and says, "Wow, they are all so close to the prediction! The sensors must be incredibly precise," it is wrong. The sensors aren't necessarily better; the system just filtered out the evidence that they weren't.

2. The "Nearest Neighbor" Problem (Picking the Best of the Best)

Sometimes, the radar sees multiple blips inside the "safe zone" (the gate). It has to pick just one to update its guess. The standard rule is Nearest-Neighbor (NN): "Pick the one that is closest to the predicted spot."

The Analogy: Imagine you have a bag of marbles, and you want to pick the one closest to the center of a target.
- If you pick one marble, it might be a little off.
- If you pick five marbles and only keep the one closest to the center, you are statistically guaranteed to pick a marble that is closer to the center than the average marble would be.
The Result: By picking the "best" match from a group, the system creates an artificial illusion of perfection. The error looks smaller than it actually is. The paper proves this is a mathematical certainty, not a bug in the code.

3. The "Self-Fulfilling Prophecy" of Tuning

This is where things get tricky for engineers. Many modern radar systems are "adaptive," meaning they learn and adjust themselves.

The Scenario: The system looks at the data (which has been filtered and picked for being "perfect") and says, "The errors are tiny! I must be overestimating the noise. I should tell the system to trust the sensors even more and ignore the noise."
The Consequence: The system lowers its estimate of how "noisy" the sensors are. It becomes overconfident.
The Paper's Finding: If you don't account for the fact that you threw away the "bad" data and picked the "best" data, your system will learn the wrong settings. It will think the sensors are perfect when they aren't.

4. The Solution: "Gate-Aware" Thinking

The paper doesn't say "stop using gates" or "stop picking the nearest neighbor." Those are necessary for the system to work. Instead, it says: Change how you interpret the numbers.

The Fix: When the system calculates its statistics, it needs to apply a "correction factor." It needs to say, "I know I threw away the big errors and picked the smallest one, so I need to mathematically 'inflate' the error numbers back to what they should be to get a true picture."
The Outcome: The paper shows that if you use this correction, the system stops learning the wrong settings. It stays close to the true noise level, rather than drifting into a false sense of security.

Summary

The paper is a warning to engineers: Just because your data looks perfect after you filter it, doesn't mean it is.

The process of filtering out "bad" data and picking the "best" match creates a statistical illusion that the system is more accurate than it really is. If you ignore this illusion, your system will tune itself incorrectly, becoming overconfident and potentially less reliable. The paper provides the exact math to fix this illusion so the system can see the truth.

Technical Summary: Selection-Induced Contraction of Innovation Statistics in Gated Kalman Filters

1. Problem Statement

In classical Kalman filtering theory, the innovation sequence (the difference between measurements and predicted states) is assumed to be zero-mean, white, and Gaussian. Under these nominal linear-Gaussian assumptions, the Normalized Innovation Squared (NIS) follows a chi-square distribution with $m$ degrees of freedom. These statistical properties serve as the foundation for filter consistency monitoring, measurement validation, and adaptive noise tuning.

However, in operational tracking systems, innovations are rarely observed in their unconditional form. Before data association and state updates, systems typically apply ellipsoidal validation gating, accepting only measurements where the NIS falls below a threshold $\tau$ . Furthermore, when multiple measurements pass the gate, Nearest-Neighbor (NN) association selects the single measurement with the minimum NIS for the update.

The core problem addressed by this paper is that these selection mechanisms—gating and NN association—fundamentally alter the statistical distribution of the observed innovations. The paper argues that standard diagnostic and tuning procedures, which rely on the assumption that observed innovations follow the nominal chi-square distribution, are systematically biased because they ignore the conditioning effects of these selection operators. Specifically, the paper investigates whether the "contraction" of innovation statistics observed in practice is a sign of filter mistuning or an inherent consequence of the selection process itself.

2. Methodology

The analysis is conducted strictly within the classical linear-Gaussian Kalman filtering framework. The authors explicitly assume that the dynamic and measurement models are correctly specified, and that process and measurement noises are zero-mean, white, and Gaussian. This isolation ensures that any observed statistical deviations are attributed solely to the structural effects of gating and association, rather than model mismatch or non-Gaussian disturbances.

The methodology proceeds in three stages:

Analytical Derivation of Gated Moments: The authors treat ellipsoidal gating as a statistical conditioning operation. They derive exact expressions for the first- and second-order moments of the innovation vector $\nu$ conditioned on the event $A = \{ \nu^\top S^{-1} \nu \leq \tau \}$ . This involves transforming the innovation into whitened coordinates to exploit rotational invariance and utilizing properties of the truncated chi-square distribution.
Order-Statistic Analysis of NN Association: The paper models NN association as an order-statistic selection mechanism. It analyzes the expected energy of the minimum-norm innovation selected from $M$ independent samples drawn from the already-gated distribution.
Monte Carlo Validation and End-to-End Simulation: The theoretical results are validated through Monte Carlo simulations. A multi-time-step constant-velocity tracking example is used to demonstrate the practical impact of these biases on adaptive measurement-noise tuning, comparing "nominal" tuning (ignoring gating effects) against "gate-aware" tuning.

3. Key Contributions and Results

A. Gate-Conditioned Innovation Statistics

The paper proves that ellipsoidal gating induces a deterministic, dimension-dependent contraction of the innovation covariance.

Mean Preservation: The conditional mean of the innovation remains zero: $E[\nu | A] = 0$ .
Covariance Contraction: The conditional covariance is a scalar multiple of the nominal covariance:
$E[\nu\nu^\top | A] = \gamma(\tau, m) S$
where $\gamma(\tau, m)$ is a contraction factor strictly between 0 and 1, dependent only on the gate threshold $\tau$ and the measurement dimension $m$ .
NIS Bias: Consequently, the expected value of the NIS for accepted measurements is $E[Z | A] = m \gamma(\tau, m)$ , which is strictly less than the nominal reference value $m$ . For example, in a 2D system with a 95% gate probability, the expected accepted NIS is approximately 1.685, not the nominal 2.0.

B. NN Association as an Additional Selection Operator

The paper demonstrates that NN association introduces a further, unavoidable contraction of innovation energy.

Order-Statistic Bias: Selecting the minimum norm among $M$ candidates from the gated distribution results in an expected energy strictly lower than the gated mean.
Impossibility Result: The authors prove that nominal innovation statistics cannot be preserved under nontrivial gating and association. Even with a perfectly matched linear-Gaussian filter, the observed innovation energy will be systematically lower than the theoretical reference due to the deterministic nature of the selection.

C. Impact on Adaptive Tuning

The end-to-end tracking study reveals the practical consequences of ignoring these effects:

Biased Tuning: Adaptive tuning schemes that use nominal chi-square references on post-gate data systematically underestimate the measurement noise scale. In the simulation, a nominal tuner learned a noise scale of ~0.80 instead of the true value of 1.0.
Gate-Aware Correction: By applying a correction factor $1/\gamma(\tau, m)$ to the observed NIS, the tuning algorithm can recover the true noise scale (learning ~1.04 in the simulation) and achieve lower Root Mean Square Error (RMSE) and Normalized Estimation Error Squared (NEES) compared to nominal tuning.

4. Significance and Claims

The paper claims that the systematic reduction in innovation energy observed in practical tracking systems is not necessarily an indicator of filter mistuning, model mismatch, or poor noise estimation. Instead, it is a fundamental, structural consequence of the selection mechanisms (gating and NN association) inherent to the tracking pipeline.

The significance of this work lies in:

Clarifying Diagnostic Interpretation: It establishes that standard NIS diagnostics must be interpreted relative to gate-conditioned reference values, not nominal ones. A "conservative" filter (low NIS) may simply be a correctly tuned filter operating under standard gating.
Preventing Self-Reinforcing Bias: It highlights that ignoring selection effects leads adaptive tuning algorithms to incorrectly reduce estimated noise levels, potentially tightening validation logic and increasing the rejection of valid measurements.
Theoretical Foundation: It provides the first exact characterization of gate-conditioned innovation moments and order-statistic biases in the context of Kalman filtering, separating structural selection effects from genuine model errors.

The authors emphasize that while the analysis is restricted to linear-Gaussian models and NN association, the identified selection effects represent a baseline bias that any nonlinear or complex association method (such as PDA or MHT) must account for, as the fundamental mechanism of conditioning on a selection event remains.