Maximum A Posteriori (MAP) Estimation

MAP estimation is fundamentally a Bayesian inference technique. It seeks the state that maximizes the posterior probability distribution, expressed via Bayes' theorem: P(state | data) ∝ P(data | state) * P(state).

• P(data | state): The likelihood function, modeling how probable the observed sensor measurements are given a hypothetical state.
• P(state): The prior distribution, encoding existing knowledge or assumptions about the state before seeing new data.
• P(state | data): The posterior distribution, the updated belief about the state after incorporating the new evidence.

The key differentiator from Maximum Likelihood Estimation (MLE) is the explicit use of a prior distribution. This allows MAP to:

• Regularize solutions, preventing overfitting to noisy data by pulling estimates toward a sensible default.
• Incorporate physical constraints (e.g., a robot's maximum velocity) or historical data directly into the estimation process.
• Handle under-constrained problems where sensor data alone is insufficient for a unique solution. The prior provides the necessary additional information to arrive at a plausible estimate.

MAP provides a single point estimate—the mode of the posterior distribution. This contrasts with other Bayesian methods (like Kalman Filters or Particle Filters) that aim to represent the entire posterior distribution (its mean, covariance, or shape).

• Advantage: Computationally efficient; yields one actionable 'best guess'.
• Limitation: Discards uncertainty information contained in the spread and shape of the full posterior. For safety-critical robotics, understanding this uncertainty is often as important as the estimate itself.

In practice, finding the MAP estimate is reformulated as a nonlinear least-squares optimization problem. This is done by taking the negative log of the posterior, turning the product of probabilities into a sum of costs:

argmin_state { -log P(data|state) - log P(state) }

• -log P(data|state): Becomes the measurement error or reprojection error.
• -log P(state): Becomes the prior error or regularization term. This sum-of-costs formulation is naturally represented and solved using a factor graph, where factors encode the likelihood and prior terms, and variables represent the unknown states.

MAP estimation is the backbone of many core robotics algorithms:

• Simultaneous Localization and Mapping (SLAM): The pose-graph optimization in modern SLAM systems is a MAP estimation problem, where poses and landmark positions are estimated given odometry and sensor constraints.
• Bundle Adjustment: In computer vision and visual odometry, bundle adjustment refines camera poses and 3D point locations by minimizing reprojection error—a MAP estimation where the prior can be weak or uniform.
• Sensor Fusion: Fusing GPS, IMU, and camera data in a graph-based optimizer (like Google Cartographer) uses MAP to find the most probable trajectory given all noisy measurements and a motion model prior.

Both are Bayesian estimators, but with different approaches and assumptions:

MAP is often used for global trajectory optimization, while Kalman filters are used for high-frequency, recursive state prediction.

Maximum A Posteriori (MAP) estimation finds the state vector x that maximizes the posterior probability P(x | z). Using Bayes' theorem, this is proportional to the likelihood P(z | x) of the measurements given the state, multiplied by the prior P(x) representing previous knowledge:

• Posterior: P(x | z) ∝ P(z | x) * P(x)
• MAP Estimate: x_MAP = argmax_x P(x | z)

This contrasts with Maximum Likelihood Estimation (MLE), which only maximizes the likelihood P(z | x), ignoring prior information. In robotics, the prior often comes from a motion model or a previous state estimate.

In practice, maximizing the posterior probability is equivalent to minimizing the negative log-posterior. This transforms the product into a sum of squared error terms, turning a probabilistic inference into a non-linear least squares optimization problem:

• Cost Function: J(x) = -log(P(z | x)) - log(P(x))
• For Gaussian models, -log(P(z | x)) becomes the Mahalanobis distance ||z - h(x)||²_Σ, where h(x) is the measurement model and Σ is the noise covariance.
• The prior term -log(P(x)) acts as a regularizer, pulling the estimate toward a prior mean. This formulation is the direct link to factor graph optimization in frameworks like g2o or GTSAM.

MAP estimation is the foundational principle behind modern graph-based Simultaneous Localization and Mapping (SLAM). The robot's trajectory (poses) and map landmarks are states in x. Constraints are derived from:

• Odometry factors (from a motion model prior).
• Landmark observation factors (from a camera or LiDAR measurement model).
• Loop closure factors (from recognizing a previously visited location).

The factor graph visually represents this factorization. Solving the SLAM problem involves performing MAP estimation over this graph, which is typically done via iterative linearization (e.g., Gauss-Newton, Levenberg-Marquardt) to find the configuration of poses and landmarks that minimizes the total error of all constraints.

Both the Kalman filter family and MAP estimation are Bayesian estimators, but they differ in scope and implementation:

• Kalman Filter (KF/EKF/UKF): A recursive filter. It maintains a running estimate and updates it sequentially as new data arrives. It produces the full posterior distribution (mean and covariance) at each step, assuming a Gaussian distribution.
• MAP Estimation (Batch/Smoothing): A batch optimizer. It considers all measurements from time 1 to T simultaneously to estimate the entire state history. It finds the mode (peak) of the posterior but does not inherently provide the full distribution (though covariance can be recovered).
• Use Case: Kalman filters are used for online, real-time filtering. MAP estimation is used for offline optimization (e.g., post-processing log data) or fixed-lag smoothing for higher accuracy.

Standard MAP assumes Gaussian noise, making it sensitive to outliers (e.g., incorrect loop closures, sensor glitches). Robust MAP estimation replaces the Gaussian measurement model with a heavy-tailed probability distribution:

• Common Robust Kernels: Huber, Cauchy, or Tukey loss functions.
• Effect: These kernels reduce the influence of large residuals, preventing a single outlier from severely distorting the entire estimate.
• Implementation: The optimization minimizes ∑ ρ(||r_i||), where ρ is the robust kernel and r_i is the residual. This is crucial for long-term autonomy where perceptual aliasing can cause erroneous data associations.

A Visual-Inertial Odometry (VIO) system like OKVIS or VINS-Fusion is a prime example of MAP estimation. The state x contains the robot's pose, velocity, and IMU sensor biases at multiple timestamps.

• IMU Factors: Provide a high-frequency motion prior between camera frames based on gyroscope and accelerometer integration.
• Visual Factors: Connect poses to observed 3D landmarks, with reprojection error based on camera geometry.
• Tightly-Coupled Fusion: The MAP estimator jointly optimizes over all IMU and visual constraints in a sliding window, marginalizing out old states to maintain computational tractability. This provides a more accurate and drift-resistant pose estimate than using either sensor alone.

Bayesian filtering is the general probabilistic framework for recursively estimating the state of a dynamic system. It applies Bayes' theorem sequentially over time, alternating between a prediction step (using a process model) and an update step (using a measurement model). MAP estimation is the specific objective of finding the maximum of the posterior distribution within this framework. Other algorithms, like the Kalman filter and particle filter, are implementations of Bayesian filtering under different assumptions (linear/Gaussian vs. nonlinear/non-Gaussian).

A factor graph is a bipartite graphical model that visually and computationally represents the factorization of a complex probability distribution. In state estimation, each factor (square node) represents a constraint from a prior, a process model, or a sensor measurement. Each variable (circle node) represents an unknown state to be estimated. MAP estimation for the entire trajectory becomes the problem of finding the variable assignments that maximize the product of all factors. This structure is the foundation for modern graph-based SLAM and bundle adjustment solvers.

These two mathematical models are the essential components of any MAP estimation problem.

• Process Model: Also called the motion or dynamics model, it describes how the system state evolves over time (e.g., x_{k+1} = f(x_k, u_k) + w_k). It encodes prior knowledge about system kinematics and is used for prediction.
• Measurement Model: Describes how sensor observations are generated from the true state (e.g., z_k = h(x_k) + v_k). It defines the likelihood function, relating what we expect to sense given a hypothetical state. The accuracy of these models directly determines the quality and physical plausibility of the MAP estimate.

The covariance matrix is a square, symmetric matrix that quantifies the uncertainty and correlations between all dimensions of an estimated state vector. In MAP estimation (and related filters like the Kalman filter), it is a central object:

• Diagonal elements represent the variance (uncertainty squared) of each state variable (e.g., position x, velocity y).
• Off-diagonal elements represent the covariance between variables (e.g., how uncertainty in orientation correlates with uncertainty in lateral position). It is used to weight residuals in the optimization, define confidence ellipsoids, and is crucial for computing the Mahalanobis distance for outlier rejection.

Since most robotic process and measurement models are nonlinear, MAP estimation typically requires solving a nonlinear least squares problem. The goal is to find the state X* that minimizes the sum of squared errors (residuals) from all constraints. This is performed using iterative solvers:

• Gauss-Newton and Levenberg-Marquardt are the most common algorithms.
• They linearize the models around the current estimate, solve a linear system (often sparse due to the factor graph), and iterate until convergence. Efficiently solving this optimization is critical for real-time performance in systems like visual-inertial odometry (VIO) and SLAM.

Maximum Likelihood Estimation (MLE) is a special case of MAP estimation where the prior term is assumed to be uniform (i.e., we have no prior knowledge about the state). Therefore, MLE finds the state that maximizes the probability of the observed data P(Z|X), ignoring P(X). MAP generalizes this by incorporating the prior P(X), which acts as a regularization term. In robotics, using a prior (e.g., a smooth motion model or a known starting point) is almost always beneficial, making MAP the more commonly applied principle.