The Optimal Guess: How the Kalman Filter Sees Through Noise

Apollo 11, July 20, 1969. The lunar module Eagle is descending toward the Sea of Tranquility at 7,000 feet. The onboard Apollo Guidance ComputerThe Block II AGC carried 2,048 words of erasable core memory and 36,864 words of fixed rope memory (16-bit words — approximately 4 KB RAM and 72 KB ROM). Source: Wikipedia, Apollo Guidance Computer. has 4 kilobytes of RAM and a processor roughly as fast as a pocket calculator from 1978. And yet it is doing something remarkable: every second, it takes noisy radar altimeter readings, fuses them with data from the inertial measurement unit, and produces the best possible estimate of the spacecraft’s altitude, velocity, and attitude. No batch processing. No data recorded and re-analyzed. Just two equations, repeated, forever.

That algorithm was 9 years old. Rudolf Kálmán, a Hungarian-American mathematician working in a Baltimore research institute, had worked it out in January 1959 and published it in March 1960 — in an engineering mechanics journal, because the controls community wasn’t convinced it would work. By the time Neil Armstrong stepped onto the lunar surface, it was running in every major aerospace navigation system on Earth.

This teardown derives the Kalman filter from scratch. Not the formula — any textbook has the formula. The derivation. Because the reason the filter works, and why it is the provably optimal estimator, is a Bayesian argument that is far more illuminating than the engineering formula it produces. Once you see why the gain K is exactly what it is, you will also understand why every nonlinear extension — the Extended Kalman Filter, the Unscented Kalman Filter, the particle filter — is a different approximation of the same underlying idea.

196019601960YEAR OF THE PAPERASME J. Basic Engineering, 82(1):35–45

4 KB4 KB4 KBAPOLLO GUIDANCE COMPUTER RAMthe filter ran on this — recursively

2222EQUATIONS PER STEPpredict + update — that's the whole algorithm

5 m5 m5 mGPS POSITION ACCURACYenabled by tightly-coupled Kalman fusion

---

The problem: every sensor lies

Start with the hard constraint. Every physical measurement is corrupted by noise. A radar altimeter measuring the lunar module’s height is accurate to within some tolerance — call it σ_v. An inertial measurement unit measuring acceleration is accurate to within some tolerance — call it σ_w. You want to know the true state of the system (altitude, velocity, attitude), but you can only observe noisy proxies.

There are two naive approaches, and both are wrong.

Trust only the sensor. Every time a measurement arrives, take it at face value. Altitude = whatever the radar says. Velocity = whatever the IMU says. The problem: sensors are noisy, and noise averages out slowly. A single bad measurement sends your estimate careening. In a low-noise environment this might be acceptable; on a lunar descent with a cramped computer, it is not.

Trust only the model. Use physics to propagate your state estimate forward: if I was at altitude 7,000 ft with velocity −20 ft/s, in 1 second I should be at 6,980 ft with a slightly updated velocity from the engine thrust. No measurements needed. The problem: model errors accumulate. Even a tiny bias in the thrust model — one percent — compounds into dozens of feet of error over a 12-minute descent. Navigation based on pure model propagation is dead reckoningestimating current position by projecting forward from a known past position using velocity and heading, without reference to external landmarks or signals. Error accumulates with time., and dead reckoning has killed ships.

Kálmán’s insight: you need both, and the right way to combine them is to weight each by how trustworthy it is. A sensor with low noise variance should be trusted heavily. A model with low process noise should be trusted heavily. The filter should lean toward whichever source has smaller uncertainty at each step. The mathematics of doing this optimally — minimizing expected squared error — produces the filter’s famous gain formula.

---

The 1D intuition: blending two Gaussians

Before any matrices, work the simplest possible case. You are estimating a scalar position x. Your model says x ≈ 50 with uncertainty 10 (meaning your prior is Gaussian with mean 50, variance 100). Your sensor reports z = 60 with noise variance 25.

What is the optimal estimate?

Two sources of information, both uncertain, both modeled as Gaussians. The product of two Gaussians is another Gaussian, and its mean is a weighted average: estimates with smaller variance get more weight. Specifically, if you combine a prior N(μ₁, σ₁²) with a likelihood N(μ₂, σ₂²), the posterior mean is:

$\mu_{post} = \mu_1 + \frac{\sigma_1^2}{\sigma_1^2 + \sigma_2^2}(\mu_2 - \mu_1)$

The fraction σ₁²/(σ₁²+σ₂²) is the Kalman gain in 1DIn one dimension with H=1: K = P_prior / (P_prior + R). When the prior is uncertain (large P_prior), K is large and the measurement dominates. When the prior is tight, K is small and the prediction dominates. — call it K. The update is simply: x̂_post = x̂_prior + K·(z − x̂_prior). And the posterior variance shrinks: σ²_post = (1−K)·σ²_prior.

Run it. The numbers make the intuition concrete:

runnable · js▸ run

output

Run the code to see its output here.

Run it. With P=100 and R=25, K=0.8 — the measurement pulls the estimate strongly because the prior is more uncertain than the sensor. Flip to P=9 and R=25 and K drops to 0.26 — the model is now tighter, so the new reading barely moves the estimate. That ratio, prior variance over total variance, is the entire intuition behind the Kalman gain. Everything else is the multivariate version of this.

---

The five equations

The full multivariable filter operates on a vector state x_k (e.g., position and velocity in three dimensions) and produces vector observations z_k (e.g., pseudoranges from four GPS satellites, or radar altitude). The system is modeled by two linear equations:

(1)

$$x_k = F x_{k-1} + w_k, \quad w_k \sim \mathcal{N}(0, Q)$$

$$x_k \in \mathbb{R}^n$$whystate vector at time k (hidden — never observed directly)

$$F \in \mathbb{R}^{n \times n}$$whystate-transition matrix — encodes your system model (e.g., constant-velocity kinematics)

$$w_k \sim \mathcal{N}(0, Q)$$whyprocess noise: unmodeled disturbances, actuation errors, wind — treated as zero-mean Gaussian

(2)

$$z_k = H x_k + v_k, \quad v_k \sim \mathcal{N}(0, R)$$

$$z_k \in \mathbb{R}^m$$whymeasurement vector (observed — what your sensor returns)

$$H \in \mathbb{R}^{m \times n}$$whyobservation matrix — the relationship between state and what you can measure

$$v_k \sim \mathcal{N}(0, R)$$whymeasurement noise: sensor inaccuracy, quantization, thermal drift

Given these two models, the Kalman filterNamed after Rudolf Emil Kálmán (1930–2016). The filter is sometimes called the Kalman-Bucy filter to credit Richard Bucy, who co-developed the continuous-time version in 1961. Source: Wikipedia, Kalman filter. runs exactly five equations per time step — two for predict, three for update:

(3)

$$\hat{x}_{k|k-1} = F \hat{x}_{k-1|k-1}$$

(4)

$$P_{k|k-1} = F P_{k-1|k-1} F^\top + Q$$

(5)

$$K_k = P_{k|k-1} H^\top \left(H P_{k|k-1} H^\top + R\right)^{-1}$$

(6)

$$\hat{x}_{k|k} = \hat{x}_{k|k-1} + K_k \left(z_k - H \hat{x}_{k|k-1}\right)$$

(7)

$$P_{k|k} = \left(I - K_k H\right) P_{k|k-1}$$

Equations (3)–(7) are the complete algorithm. ¹1In practice, the numerically stable “Joseph form” of the covariance update — P = (I−KH)P(I−KH)ᵀ + KRKᵀ — is preferred over equation (7) because floating-point rounding can make the naive form non-positive-definite. The result is mathematically identical; the form matters only for 64-bit precision. MDPI Sensors 2023 The predict step runs whenever time advances (even with no measurement). The update step runs whenever a measurement arrives. On a GPS receiver, predict runs at 100 Hz from the IMU, and update runs at 1–10 Hz when satellite pseudoranges are available.

---

The architecture

Here is what those five equations look like as a data-flow system. The diagram makes one thing visible that the equations obscure: the predict-update cycle is a feedback loop. The posterior at step k is the prior at step k+1.

Nodes

• Prior state (x̂_{k-1}, P_{k-1})
• PREDICT Eqs. (3)–(4)
• Sensor measurement z_k
• Kalman gain K Eq. (5)
• UPDATE Eqs. (6)–(7)
• Posterior state (x̂_k, P_k)

Connections

• Prior state (x̂_{k-1}, P_{k-1}) → PREDICT Eqs. (3)–(4) (advance time)
• PREDICT Eqs. (3)–(4) → Kalman gain K Eq. (5) (P_{k|k-1})
• PREDICT Eqs. (3)–(4) → UPDATE Eqs. (6)–(7) (x̂_{k|k-1})
• Sensor measurement z_k → UPDATE Eqs. (6)–(7) (innovation ỹ = z − Hx̂)
• Kalman gain K Eq. (5) → UPDATE Eqs. (6)–(7) (weight K)
• UPDATE Eqs. (6)–(7) → Posterior state (x̂_k, P_k) (posterior)
• Posterior state (x̂_k, P_k) → Prior state (x̂_{k-1}, P_{k-1}) (feeds next step k→k+1)

The cycle arrows matter. After the update at step $k$, the posterior $(\hat{x}_{k|k},\ P_{k|k})$ becomes the input to the predict step at $k+1$. The filter never re-processes old measurements — it just carries their effect forward through the covariance $P$. That is what makes it recursive: the entire history of observations is summarized in the current $(\hat{x},\ P)$ pair, and nothing else is needed.

This is also why the filter ran on the Apollo Guidance Computer’s 4 kilobytes of RAM: it does not accumulate measurements; it consumes them one by one, discards them, and moves on. A batch estimator would have needed to store every radar ping since engine ignition. The Kalman filter needed only the current state vector and covariance matrix — a handful of numbers no matter how long the mission had been running.

---

Why K is exactly what it is: the Bayesian proof

The gain formula (5) looks arbitrary unless you derive it. The derivation is short, and it is worth doing once — it converts the filter from a formula you memorize into a result you understand.

The setup: at the predict step, you have a Gaussian prior over the state:

$p(x_k | z_1, \ldots, z_{k-1}) = \mathcal{N}(\hat{x}_{k|k-1},\ P_{k|k-1})$

A measurement z_k arrives. By the observation model, the likelihood of z_k given x_k is also Gaussian:

$p(z_k | x_k) = \mathcal{N}(H x_k,\ R)$

The posterior is proportional to prior × likelihood. Both are Gaussian, and the product of Gaussians is Gaussian. The posterior mean minimizes the expected squared error — it is the MAP estimatorMaximum A Posteriori: the estimate that maximizes the posterior distribution p(x|z). For a Gaussian posterior, the MAP equals the mean and minimizes expected squared error (MMSE). and the MMSE estimate simultaneously.

To find it, complete the square in the log-posterior:

(8)

$$\ln p(x_k | z_k) \propto -\frac{1}{2}\left[(z_k - Hx_k)^\top R^{-1}(z_k - Hx_k) + (x_k - \hat{x}_{k|k-1})^\top P_{k|k-1}^{-1}(x_k - \hat{x}_{k|k-1})\right]$$

$$\text{likelihood term: } (z_k - Hx_k)^\top R^{-1}(z_k - Hx_k)$$whypenalizes states that predict a measurement far from what we observed

$$\text{prior term: } (x_k - \hat{x}_{k|k-1})^\top P_{k|k-1}^{-1}(x_k - \hat{x}_{k|k-1})$$whypenalizes states far from the predicted mean — weighted by prior precision

Expanding, collecting x_k terms, and completing the square yields the posterior mean and covariance:

(9)

$$\hat{x}_{k|k} = \hat{x}_{k|k-1} + P_{k|k-1} H^\top \underbrace{\left(H P_{k|k-1} H^\top + R\right)^{-1}}_{= S_k^{-1}} (z_k - H\hat{x}_{k|k-1})$$

$$K_k = P_{k|k-1} H^\top S_k^{-1}$$whyK is the ratio: prior covariance (mapped to measurement space) over total covariance

$$P_{k|k} = \left(P_{k|k-1}^{-1} + H^\top R^{-1} H\right)^{-1}$$whyposterior precision = prior precision + measurement precision — information adds

$$P_{k|k} = (I - K_k H) P_{k|k-1}$$whyequivalent form (via Woodbury identity) — this is equation (7)

The gain $K_k = P_{k|k-1} H^\top (H P_{k|k-1} H^\top + R)^{-1}$ is not an arbitrary engineering choice. It is the exact coefficient that minimizes the posterior uncertainty — derived directly from Bayes’ theorem applied to two Gaussian distributions. This result has a name: the Kalman filter is the minimum mean-square error (MMSE) estimator for linear-Gaussian systemsKálmán (1960) Theorem 8 proved that the filter minimizes E[(x̂−x)²] among all linear estimators. For Gaussian noise, it also minimizes among ALL estimators, linear or not. Source: Wikipedia Kalman filter; Kálmán 1960 ASME §4.. When noise is Gaussian, no other filter — linear or nonlinear — can do better.

The 1D formula from the intuition section is exactly this with H=1: K = P/(P+R). The derivation confirms it was right all along.

Go deeper: the information form and why precision adds

The posterior covariance P = (P⁻¹ + Hᵀ R⁻¹ H)⁻¹ shows that information (inverse covariance) adds linearly. Each measurement contributes Hᵀ R⁻¹ H bits of information. This is the “information filter” form of the update — equivalent to equation (7) but numerically different. When you have many sensors feeding in at once (like 12 GPS satellites), the information filter computes the combined update by summing their contributions, then inverting once. The standard form (equation 7) is equivalent but applies updates sequentially. For the Kalman-Bucy continuous-time filter (1961), the Riccati differential equation dP/dt = FP + PFᵀ + Q − PHᵀ R⁻¹ HP takes this additive-information form naturally.

---

Running the full filter

One equation at a time is clear. Running it on actual data reveals the behavior: how the estimate smooths out measurement noise, how the covariance converges, how changing Q and R shifts the filter’s personality.

runnable · js▸ run

output

Run the code to see its output here.

Run it with the defaults (Q=1, R=9) and the filter reduces RMS noise by roughly 40–50%. Now set Q=0.1 and R=9: with a tight process model, the filter trusts its own prediction heavily and barely moves when measurements arrive — smooth but slow to respond to genuine state changes. Set Q=10 and R=1: with an uncertain model and a precise sensor, K approaches 1 and the estimate tracks the measurements closely, accepting their noise. Q and R are the knobs that encode your engineering judgment about which source to trust. The mathematics then does the optimal blending.

---

The Riccati equation and convergence

There is a deep result buried in equations (4) and (7): under constant F, H, Q, R, and mild technical conditions (the system must be observableA system is observable if you can infer the complete state from a finite history of measurements. Formally, the observability matrix has full rank. An unobservable state dimension cannot be corrected by any measurements, so the filter's uncertainty in that dimension grows without bound.), the covariance P converges to a fixed steady-state value P_∞ regardless of the initial P_0.

The update equation $P_{k+1} = F (I-KH) P_k F^\top + Q$ is a discrete Riccati equationA nonlinear matrix difference (or differential) equation of the form P_{k+1} = f(P_k). Named after Jacopo Riccati (1724). Appears in optimal control (LQR) as well as Kalman filtering, because they are mathematical duals — a connection Kálmán himself noted in his 1960 paper.. It contracts: P_∞ is the fixed point. In 1D, P_∞ satisfies P_∞ = R·P_∞/(P_∞+R) + Q, which gives P_∞ = (Q + √(Q²+4QR))/2. The corresponding steady-state gain K_∞ = P_∞/(P_∞+R) can be precomputed offline and hardcoded into the embedded system — no matrix inversion required at runtime.

chart: line of gain by step — 10 points

The steep drop in the first two steps is typical: the filter rapidly learns from early measurements and the gain plummets toward its asymptote. After step 5, the gain is essentially constant. This is why embedded Kalman implementations often skip the covariance update entirely and use a precomputed fixed gain — valid for linear, stationary systems with stable Q and R.

The second chart shows how the steady-state gain responds to the ratio of process noise to measurement noise. When your model is noisy relative to your sensor (high Q/R), the filter rightly trusts the measurements more. When the model is tight relative to the sensor (low Q/R), the predictions dominate.

chart: line of gain by qr — 6 points

The golden ratio at Q/R=1 is not a coincidence — it falls out of the quadratic formula for P_∞ when Q=R, and 1/φ is the positive root of K² + K − 1 = 0. ²2The Kalman filter and the Linear Quadratic Regulator (LQR) optimal controller are mathematical duals. The Riccati equation for the filter covariance is identical in structure to the Riccati equation for the LQR cost matrix — with time running backwards. Kálmán recognized this duality in his 1960 paper and used it in the proof. It means designing an optimal estimator and designing an optimal controller are literally the same computation. Wikipedia

---

1960: the paper and the Moon

Rudolf Emil Kálmán was born in Budapest, Hungary, on May 19, 1930Source: Wikipedia, Rudolf E. Kálmán; MacTutor History of Mathematics, University of St Andrews biography of Kalman.. He emigrated to the United States, earned his bachelor’s and master’s degrees from MIT in 1953 and 1954, and completed his doctorate at Columbia University in 1957³3MacTutor (University of St Andrews) records that Kálmán submitted his thesis and was awarded the D.Sci. in 1958; Wikipedia gives 1957. The discrepancy likely reflects thesis submission versus degree conferral date. MacTutor. By 1958, he was a research mathematician at the Research Institute for Advanced Studies (RIAS) in Baltimore — not a controls lab, but a theoretical institution that gave him the freedom to think broadly.

He worked out the core idea in January 1959 and presented it to the American Society of Mechanical Engineers in March of that year. The paper was received by the ASME on February 24, 1959 and published in the ASME Journal of Basic EngineeringVol. 82, No. 1, 1960, pp. 35–45. The paper appeared in a mechanical engineering journal rather than an electrical engineering or controls venue because the controls community was skeptical. Source: Kálmán (1960) DOI:10.1115/1.3662552; georgiou.eng.uci.edu/publications/asme.pdf (IEEE/UCI memorial). on March 1, 1960. Title: “A New Approach to Linear Filtering and Prediction Problems.”

The controls community was not immediately convinced. The Wiener filter — which solved the same problem in the frequency domain, for stationary signals, using batch processing — had dominated signal processing since the 1940s. Kálmán’s paper claimed a recursive solution that worked for non-stationary signals without storing any measurement history. It seemed too good to be true.

The reception changed in fall 1960. Kálmán visited Stanley F. Schmidt at NASA Ames Research CenterMountain View, California. Schmidt was chief of the Dynamic Analysis Branch (per NASA Aeronautics), leading a team of eight engineers working on circumlunar navigation for the Apollo program. Source: NASA Aeronautics ('Math Invented for Moon Landing'); Bhatta (2023), citing McGee-Schmidt NASA Tech Memo 1985 and Schmidt oral history 2014; IEEE aerospace history paper (ieeexplore.ieee.org/document/5466132/). in Mountain View, California. Schmidt’s team was wrestling with a navigation problem the Wiener filter could not solve: fusing radar altimeter readings and IMU data on a spacecraft flying a non-stationary trajectory, on a computer with kilobytes of memory. Stationary signals, batch processing, frequency domain — none of that fit.

They all sat there, listening to Rudy, and I did, too. When he got through, I don’t think any of them understood, and I didn’t understand much, but after his discussion of it, I said, ‘I think this might have something to do with the [navigation problem].

Schmidt’s team spent the next several months adapting the filter for the nonlinear case — linearizing the dynamics around the current trajectory estimate — and by early 1961, they had validated it through simulations. The technique had been extended for nonlinear problems even before the original linear paper had circulated widely. The Extended Kalman Filter was loaded onto Apollo 11Source: Aerospace America / AIAA 'Honoring a Legacy Algorithm'; IEEE Kalman Filtering Aerospace Applications 1960–present (ieeexplore.ieee.org/document/5466132/). — the Apollo Guidance Computer used it to estimate position, velocity, and attitude from radar measurements and inertial sensors throughout the lunar descent and ascent.

The continuous-time extension, the Kalman-Bucy filterKálmán, R.E. & Bucy, R.S. (1961). 'New Results in Linear Filtering and Prediction Theory.' ASME Journal of Basic Engineering, 83(1):95–108. Extended the 1960 discrete-time filter to continuous-time stochastic differential equations., followed in 1961, coauthored with Richard Bucy. The filter now had both a discrete form (useful for digital computers sampling at fixed intervals) and a continuous form (expressed as a stochastic differential equation, useful for analog systems and theoretical analysis). Kálmán later received the National Medal of Science — presented by President Obama in 2009 — the Charles Stark Draper Prize, and the IEEE Medal of Honor for work that touches every navigation system built since 1960.

---

What breaks: the linear-Gaussian assumption

The filter’s optimality proof rests on two assumptions: that the state evolution and observation are linear (F and H are constant matrices, not functions of x_k) and that the noise is Gaussian. Break either and the proof fails. Different extensions break it in different ways — and each is a different approximation of the same underlying Bayesian update.

1. 4Particle Filter (Sequential Monte Carlo)

   Arbitrary nonlinearity, arbitrary noise distributions. Represents the posterior as a weighted set of random 'particles' drawn from a proposal distribution. Asymptotically exact as particle count → ∞. Cost: O(N_particles) per step; degrades in high dimensions (curse of dimensionality).

   Works for any system — but computational cost explodes with state dimension. Used in robot localization, radar tracking, financial filtering.

2. 3Unscented Kalman Filter (UKF)

   Nonlinear systems, Gaussian-assumed noise. Uses a deterministic set of 2n+1 sigma points to propagate mean and covariance through the nonlinearity exactly to second order — without computing Jacobians. More accurate than EKF for moderate nonlinearity.

   Julier & Uhlmann (1997). Sigma points capture the mean and covariance exactly; after propagation through f(x), the resulting mean and covariance are accurate to 3rd order (2nd order for Gaussian noise). No Jacobian required.

3. 2Extended Kalman Filter (EKF)

   Nonlinear systems, Gaussian-assumed noise. Linearizes f(x) and h(x) via first-order Taylor expansion around the current estimate. Requires computing the Jacobian matrices F_k = ∂f/∂x and H_k = ∂h/∂x at each step.

   Schmidt's adaptation for Apollo (1961). Works well when the system is 'nearly linear' on the timescale of a single step. Can diverge for strongly nonlinear systems — the linearization error is not bounded.

4. 1Kalman Filter (linear-Gaussian)

   Linear state evolution and observation, Gaussian noise. EXACT: the posterior is Gaussian, and equations (3)–(7) compute it exactly. Minimum mean-square error, provably optimal. Source: Kálmán (1960), Theorem 8.

   The only case where the Bayesian posterior is analytically tractable and computable in O(n³) time. Every other filter is approximating what this one computes exactly.

The layering is the argument: each filter above is trying to compute the same Bayesian posterior, p(x_k | z_1, …, z_k). In the linear-Gaussian case, that posterior is exactly Gaussian, and the Kalman filter computes it exactly. Break linearity (EKF, UKF) and you have to approximate the propagation of the Gaussian through a nonlinear function. Break Gaussianity (particle filter) and you abandon the parametric form entirely and represent the posterior with samples.

1. causeSystem dynamics are nonlinear: x_{k+1} = f(x_k) + w_k, f ≠ Fx_k

   The predict step P_{k|k-1} = F P F^T + Q requires a linear F. With a nonlinear f, the Gaussian prior is mapped to a non-Gaussian predicted distribution — the Kalman formula for the predicted covariance is no longer exact.

2. therefore
   effectEKF: linearize f around the current estimate using the Jacobian

   F_k = ∂f/∂x|_{x̂} turns the nonlinear update into an approximately linear one. Works when x stays near the linearization point, but the first-order Taylor error grows with the curvature of f.

3. therefore
   effectEKF divergence: when linearization error dominates, estimates become overconfident and ignore measurements

   If the true state departs from the linearization trajectory by more than the noise level, the filter's covariance underestimates true uncertainty. K drops, measurements are down-weighted, and the estimate locks onto a wrong value.

4. and so
   turning pointUKF: bypass Jacobians with sigma points — more accurate, same cost

   Choose 2n+1 deterministic points that match the Gaussian's mean and covariance exactly. Propagate each through f(x). The output mean and covariance capture nonlinear effects to 3rd order.

5. which drives
   causeNoise is non-Gaussian: impulsive, multimodal, or heavy-tailed

   GPS multipath creates bimodal pseudorange errors. Radar clutter creates impulsive spikes. A single outlier measurement with a Gaussian assumption will corrupt the estimate for many subsequent steps.

6. therefore
   effectParticle filter: abandon Gaussian assumption, represent posterior with N weighted samples

   Any noise distribution, any nonlinearity. At N=1,000 particles in 3D, the posterior is well-represented. But N must scale exponentially with state dimension — a 10D state needs millions of particles for coverage.

The practical message: reach for the Kalman filter when your system is linear or nearly linear and your noise is roughly Gaussian. Use the EKF for mildly nonlinear systems where computing Jacobians is feasible. Use the UKF when computing Jacobians is painful or the nonlinearity is stronger. Use the particle filter when the posterior is multimodal, when noise is impulsive, or when the state dimension is small enough to afford it.

---

Where it runs today

The filter Kálmán published in a mechanical engineering journal in 1960 now runs in approximately every serious navigation system on Earth. The applications share a structure: hidden state, noisy measurements, Gaussian-ish noise, need for real-time estimates.

GPS/IMU sensor fusion. The companion primer on how GPS actually locates you showed how four satellite pseudoranges give a position fix via least-squares. What it deferred: between satellite measurement updates (available at 1–10 Hz), the receiver’s IMU accelerometers and gyroscopes propagate the state at 100–400 Hz. A tightly coupled Kalman filterIn tightly coupled integration, the Kalman filter's state vector includes receiver position, velocity, clock bias, and IMU biases. Measurements are raw pseudoranges from individual satellites (not position fixes from a standalone GNSS solver). This enables operation with fewer than 4 satellites and directly models measurement noise. Source: MDPI Sensors 2023 (mdpi.com/1424-8220/23/7/3676). fuses these directly: state vector = [position, velocity, clock bias, IMU biases], measurement vector = raw pseudoranges. When the satellite count drops below four, the IMU-propagated prediction keeps the estimate alive. In urban canyons where multipath corrupts pseudoranges, the filter downweights those measurements through inflated R entries.

Rocket landing. SpaceX Falcon 9 first-stage landings use state estimation throughout the powered descent. The dynamics are highly nonlinear (thrust vector control, varying mass as propellant burns, aerodynamic forces at low altitude) — this is firmly EKF or UKF territory. The state vector includes position, velocity, attitude quaternion, thrust direction, and fuel mass. Measurements come from radar altimetry, GPS, and onboard IMU. The filter must run reliably under vibration, aerodynamic perturbation, and engine gimbaling. A divergent estimator means a missed landing.

Quantitative finance. Asset prices are not the state — they are noisy observations of a hidden underlying process (fair value, momentum regime, volatility). A financial state-space modelObservation equation: y_t = Z x_t + noise. State equation: x_{t+1} = T x_t + noise. The Kalman filter estimates the hidden state x_t (e.g., trend, volatility, hedge ratio) from noisy prices y_t. Used in pairs trading, volatility forecasting, and yield curve modeling. Source: Quant Analysis (quantanalysis.org.uk/financial-markets/kalman-filter). models the price as an observation of a hidden state (trend + noise), and the Kalman filter tracks that hidden state recursively, adapting as new prices arrive. The key advantage over a rolling window: the filter weights older data naturally through the Riccati equation, without a hard cutoff.

Weather and climate modeling. The Ensemble Kalman Filter (EnKF)A Monte Carlo variant of the Kalman filter for very high-dimensional systems (e.g., numerical weather prediction with millions of state variables). Approximates the covariance P with an ensemble of N model runs (typically N=20–100), keeping the computational cost tractable. Widely used in operational numerical weather prediction. is the backbone of modern numerical weather prediction. The state is the global atmosphere — on the order of 10⁷ variables. The full covariance matrix would be 10¹⁴ entries, computationally impossible. The EnKF approximates it with an ensemble of ~50–100 perturbed model runs, and updates all of them when observations (radiosonde data, satellite retrievals, surface stations) arrive.

---

Failure modes

Every teardown owes you the edges. The filter is optimal inside its assumptions. Here is what happens at those edges:

filter5 rows

Kalman filter failure modes, root causes, and mitigations.

Failure mode↕	Root cause↕	Observable symptom↕	Mitigation↕
Filter divergence	Q is set too low — filter believes its model is perfect and stops incorporating measurements	Gain K → 0; estimate locks onto wrong value; innovation grows unbounded	Increase Q to reflect real model uncertainty. Use adaptive noise estimation or innovation monitoring.
EKF linearization failure	System is strongly nonlinear; first-order Taylor approximation introduces large errors	Covariance collapses (overconfident) while true error grows; catastrophic divergence on highly curved trajectories	Switch to UKF (sigma points) or particle filter. Reduce step size so linearization is valid over each step.
Non-Gaussian outliers	Sensor returns occasional large-error measurements (GPS multipath, radar clutter) not modeled by Gaussian R	A single bad measurement corrupts estimate for many steps; RMS error spikes	Robust filtering: chi-squared innovation test to gate outliers; Huber loss instead of squared error; particle filter for non-Gaussian R.
Unobservable state dimensions	Observation matrix H does not span the full state space — some dimensions are invisible to measurements	P grows without bound in unobservable directions; those state dimensions drift freely	Add sensors to cover unobservable modes, or remove unobservable dimensions from the state vector. Check the observability Gramian.
Computational: P matrix O(n³) cost	Full covariance update requires n×n matrix multiplication and inversion — cubic in state dimension	Intractable for large state vectors (e.g., 10⁷-dimensional weather model)	Information filter form; ensemble Kalman filter (low-rank P approximation); covariance inflation + localization for geophysical models.

The most common failure in practice is the first: underspecified process noise Q. Engineers who are proud of their model set Q to zero or near-zero. The filter then places infinite confidence in its own predictions and the gain drops toward zero — measurements stop being incorporated. Any real disturbance sends the estimate irrecoverably astray. The fix is counter-intuitive: admit that your model is wrong, specify Q generously to reflect genuine model uncertainty, and let the filter do its job of weighting the model against the sensor.

---

Check your understanding

1. Q1The Kalman gain K = P·Hᵀ·(H·P·Hᵀ + R)⁻¹ can be interpreted as a ratio. What ratio?

   • Sensor precision divided by model precision
   • Prior state uncertainty (projected into measurement space) divided by total uncertainty in measurement space
   • The number of measurements divided by the state dimension
   • The ratio of process noise Q to measurement noise R

2. Q2Why is the Kalman filter recursive — why does it not need to store all past measurements?

   • It assumes measurements are independent so the past is irrelevant
   • The current (x̂, P) pair is a sufficient statistic for all past observations under the Gaussian model — the full posterior is captured in two numbers (vector and matrix)
   • It compresses measurements using run-length encoding
   • Older measurements are assumed to be less accurate

3. Q3The Extended Kalman Filter linearizes the nonlinear dynamics around the current state estimate. What failure mode does this cause?

   • The filter becomes too computationally expensive to run in real time
   • The covariance update requires storing the full measurement history
   • Linearization error can make the filter overconfident — covariance P collapses while true error grows, and subsequent measurements are ignored
   • The filter cannot handle multivariate states

---

The mental model to take away

Three ideas, one algorithm.

Predict with a model. Update with a measurement. Weight by uncertainty. This is the complete description. The model propagates your best guess forward and inflates your uncertainty (P grows). The measurement arrives, and you fold it in — weighted by how much of the total uncertainty it explains (K). The posterior is tighter than either input alone, because two independent sources of information about the same quantity always reduce uncertainty.

The gain K is the Bayesian posterior in closed form. Equations (5)–(7) are not engineering heuristics. They fall out of completing the square in the log-posterior of two Gaussians. Kálmán proved in 1960 that this derivation produces the minimum mean-square error estimator among all linear filters — and under Gaussian noise, among all filters. When engineers say the Kalman filter is “optimal,” this is what they mean: there is a proof, not just a track record.

Every nonlinear extension is an approximation of the same computation. The EKF approximates the Gaussian-to-Gaussian propagation with a first-order Taylor expansion. The UKF approximates it with a carefully chosen set of sigma points accurate to third order. The particle filter abandons the Gaussian assumption entirely and represents the posterior as a cloud of weighted samples. Each is trading computational tractability for exactness — and each is trying to compute the same thing: the Bayesian posterior update that the Kalman filter computes exactly when the linear-Gaussian assumption holds.

The next time you see a GPS position fix update smoothly rather than jittering between noisy satellite readings, or watch a Falcon 9 land on a drone ship in 30-knot winds, or see a quant model track a time-varying spread — somewhere in that system, a pair of matrix equations is running every few milliseconds, asking the same question Kálmán worked out in Baltimore in January 1959: given everything I knew before, and this one new noisy measurement, what is the best estimate I can make now?

The answer has not changed in 65 years.