Where is it, when every sensor only knows a direction
An acoustic array, a camera, a direction finder: each gives a bearing, none gives a position. Two of them give a fix whose quality is set by geometry, not by their noise. Drag the sensors and the target, and watch the error ellipse tell you when the answer is a point and when it is a line.
Reading the ellipse
Each sensor constrains the target to a thin wedge around its bearing. Where two wedges cross, the target must be. If they cross at a right angle the intersection is a small diamond; if they cross at a few degrees it is a long sliver along the lines of sight. The ellipse is that sliver computed exactly: the inverse of the information matrix J = Σ (1 / (σ² r²)) · u uᵀ, one term per sensor, where u is the unit vector perpendicular to the bearing and r the range. It is the best any unbiased estimator can do.
Two things follow. Range hurts quadratically: doubling the distance quadruples the uncertainty for the same bearing noise. And a target on the line through two sensors is unobservable: both bearings are the same line, the information matrix has a zero eigenvalue, and the ellipse is infinitely long. No amount of averaging fixes geometry; only a third sensor off that line does. Press the button to see it happen.
The Monte Carlo dots are honest fixes: noisy bearings, a pseudo-linear least-squares first guess, then a few Gauss-Newton iterations of the maximum-likelihood cost. At small noise they fill the ellipse. At large noise the first guess alone is biased toward the sensors, which is why the reference implementation refines it: the results page reports the size of that bias.
Position error (root of the trace of the bound) against the angle at which the two bearings cross, at the ranges and bearing noise currently on the map. Dots: experiment E4, measured with noisy bearings, when docs/data/e4.json exists.
Fusion is the same machinery one level up
Multi-sensor tracking is not a separate algorithm. In the reference tracker each sensor's measurement model (bearing, range-and-bearing, position) is applied in turn within a scan: predict once, then update with each sensor's detections through its own linearised model. The information matrix on this page is exactly the sum the Kalman filter accumulates, one sensor at a time. What fusion buys is visible here: the third sensor does not average the error down, it removes the direction in which the first two had none.
tracklab.fusion holds triangulate, gdop and observable; the map runs the same formulas.