Signal geolocation accuracy: CEP, error budgets and sensor geometry

A geolocation fix is not a point — it is a probability distribution. The coordinate an analyst sees on the common operating picture is the peak of that distribution; the question the analyst almost never gets answered explicitly is how wide the distribution is and in which direction. Two systems can report exactly the same fix coordinate for exactly the same emitter and yet carry completely different operational meaning: one may have a true 90% confidence radius of 80 meters; the other 4 kilometers. Unless the consumer knows which is which, they are flying blind. This article builds the vocabulary and arithmetic to answer that question rigorously: how accuracy is defined and measured, where the error comes from in both TDOA and AOA systems, how geometry amplifies or compresses measurement noise into position error, and how all of it should be communicated to the people who act on it.

Geolocation accuracy metrics: CEP, CEP50, CEP90, and the error ellipse

The most common accuracy metric in defense geolocation is the circular error probable, or CEP. CEP50 — the subscript is almost always dropped in conversation — is the radius of the smallest circle centered on the true emitter position that contains 50% of independently repeated fix estimates. The definition is purely statistical: if you ran the same collection-and-solution scenario a large number of times under the same conditions, half the resulting coordinates would fall inside the CEP circle and half outside. The number says nothing about which side of the circle a specific fix lands on.

CEP90 extends the same idea to the 90th percentile: the circle containing 90% of fixes. For many mission applications, CEP90 is the operationally relevant figure. Targeting, for example, is driven by the worst case that 90% of fixes will not exceed, not by the median. Intelligence consumers specifying accuracy requirements should state which percentile they need; engineers designing the system should budget to CEP90 with margin.

For a bivariate-normal position error with equal variances σ in both horizontal directions — the circular case — the conversion between percentiles follows fixed multipliers:

CEP50  = 1.1774 · σ         (radius enclosing 50% probability)
CEP90  = 2.1460 · σ  ≈ 1.82 · CEP50
CEP95  = 2.4477 · σ  ≈ 2.08 · CEP50
CEP99  = 3.0349 · σ  ≈ 2.58 · CEP50

These multipliers assume circularity. Real geolocation error distributions are almost never circular — sensor geometry, terrain masking, and the directional sensitivity of different measurement types all impose an elliptical shape on the error region. The correct representation is the error ellipse: a 2D confidence region with semi-major axis a, semi-minor axis b, and orientation angle θ relative to north. The error ellipse comes directly from the covariance matrix of the position estimate: its axes are the eigenvalues of the 2×2 horizontal position covariance, and the eigenvectors give the orientation.

Operationally, when do you report CEP versus the full ellipse? CEP is the right choice when a single summary number is required for comparison, status reporting, or a specification. The error ellipse is necessary when the directional character of uncertainty matters — for example, when the emitter is near the boundary of an area of interest and the question is whether it is inside or outside, or when the elongated axis of the ellipse happens to align with a critical direction. As a rule of thumb, report the full ellipse whenever the axis ratio a/b exceeds 1.5.

TDOA error budget: the dominant contributors

A TDOA error budget decomposes the total measurement uncertainty into its contributing sources, expressed in common units of time (nanoseconds) or range difference (meters, after multiplying by the speed of light). The budget then feeds the GDOP calculation to predict the resulting position error. Four sources dominate in practice.

Timing synchronization error is usually the largest single contributor. TDOA is built on measuring the difference in signal arrival time between two sensors; that difference is computed by comparing timestamps attached to the received signals at each site. If the two clocks disagree about what time it is, a bias is injected directly into the TDOA measurement. GPS-disciplined oscillators hold inter-sensor clock agreement to approximately 20–100 ns (6–30 m of range-difference error). Fiber-distributed timing using Precision Time Protocol (PTP) or White Rabbit protocols achieves sub-nanosecond synchronization. The residual after GPS discipline is not constant — it drifts with temperature and oscillator aging — so a robust system estimates the inter-sensor clock offset as a nuisance parameter in the solver, using a known reference emitter for calibration when available.

Cross-correlation measurement noise sets the floor on how precisely the delay can be estimated from the signal itself. For a signal of bandwidth B (Hz) at signal-to-noise ratio SNR, the standard deviation of the TDOA estimate from a single cross-correlation is approximately:

σ_τ  ≈  1 / (2π · B · √SNR)     [seconds]
σ_r  =  c · σ_τ                  [meters of range difference]

Example:  B = 1 MHz,  SNR = 20 dB (power ratio 100)
  σ_τ  ≈  1 / (2π · 1×10⁶ · 10)  ≈  16 ns
  σ_r  ≈  4.8 m

Narrowband signals — an HF voice link at 3 kHz, or a frequency-hopping burst with a short dwell — produce much larger σ_τ. Wideband emitters are inherently easier to locate precisely for exactly this reason. Integration time also matters: coherently integrating for longer effectively increases the processing gain, sharpening the cross-correlation peak and reducing σ_τ for a given instantaneous SNR.

Multipath bias is structurally different from the two noise sources above: it is a bias, not random noise. A reflected copy of the emitter signal arrives at the sensor after the direct path, delayed by the additional path length through the reflection. If the reflection is strong enough, the cross-correlation peak shifts toward the spurious path or develops a shoulder that biases the peak-picking algorithm toward a longer apparent delay. Unlike noise, multipath cannot be averaged away by integrating longer. Mitigations include pre-correlation whitening to reduce the sidelobe structure, super-resolution algorithms that attempt to resolve the direct and reflected paths, and spatial filtering (beam steering toward the emitter direction at each sensor). In urban and mountainous terrain, multipath is often the practical accuracy ceiling.

Ionospheric group delay applies specifically to HF signals (3–30 MHz) that propagate via skywave. The ionosphere is a dispersive medium: its refractive index varies with frequency according to the Appleton-Hartree equation, so different spectral components of a wideband HF signal travel different effective path lengths. More critically for TDOA, the group delay on the same sky-wave path differs between two sensors not only because they are at different ranges from the emitter but also because the ionospheric electron density profile along the path to each sensor differs. Unmodeled differential ionospheric delay easily introduces 5–50 μs of TDOA bias (1.5–15 km of range-difference error). Mitigation requires either an ionospheric model (IRI, with TEC map corrections from GNSS receivers) applied to each path, or calibration against a reference emitter whose location is known, whose signals propagate on a similar ionospheric path.

Contributor	Typical range	Type	Primary mitigation
Clock synchronization error	20–100 ns (GPS); <1 ns (fiber PTP)	Bias + drift	Fiber PTP; nuisance estimation in solver
Cross-correlation noise	1–50 ns depending on B and SNR	Random	Wider bandwidth; longer integration
Multipath bias	10–500 ns (environment-dependent)	Bias	Whitening; super-resolution; spatial filtering
Ionospheric delay (HF only)	5–50 μs differential	Bias	IRI/TEC model correction; reference-emitter calibration

AOA (bearing) error budget

Angle of arrival direction finding measures the bearing from the sensor to the emitter. Each bearing intersects the emitter location somewhere along a ray; multiple bearings from different sensors — or from one sensor at different times if the emitter or sensor moves — triangulate a fix. The error budget for an AOA system differs fundamentally from TDOA: the dominant noise sources are in the antenna and analog signal chain, not in the synchronization infrastructure.

Antenna calibration error is the fixed offset between the array's mechanical and electrical boresight. A poorly calibrated array has a systematic bearing error that is constant across all measurements from that sensor. Calibration is typically done by observing a reference emitter at a precisely known location and comparing the measured bearing to the geometric truth. Well-calibrated tactical arrays achieve boresight errors of 0.5–2°; poorly calibrated ones can be off by 5° or more. Since calibration error is systematic, it introduces a bias into every bearing measurement from that sensor — it does not average away with more observations.

Mutual coupling between array elements distorts the phase relationships that direction-finding algorithms rely on. When element currents interact through electromagnetic coupling, the effective array manifold deviates from the theoretical free-space pattern used in algorithm calibration. MUSIC, ESPRIT, and beamforming algorithms are all sensitive to manifold mismatch: even small coupling-induced phase errors can shift the estimated bearing by fractions of a degree, and the effect is direction-dependent. The correction is to measure the actual array manifold by rotating the array in an anechoic environment, or by fitting a coupling matrix model to a set of calibration measurements. For fixed installations, this is practical; for rapidly deployed tactical sensors, the coupling environment may shift significantly with proximity to vehicles, buildings, or terrain.

Pattern distortion arises from the ground plane, nearby reflectors, and the structure on which the antenna array is mounted. These perturb the effective phase center of each element in a direction-dependent way, introducing bearing errors that vary with azimuth. Pattern distortion is generally modeled or measured in the same calibration sweep as mutual coupling, and both are folded into a corrected array manifold stored in the sensor's database.

Baseline length and aperture govern the fundamental angular resolution. For a linear array of length L at wavelength λ, the Rayleigh resolution limit is approximately λ/L radians. For a 10-element UHF array at 400 MHz (λ = 0.75 m) spanning 4 m, this is about 0.75/4 ≈ 0.19 rad (10.7°). Super-resolution algorithms like MUSIC can resolve below this limit at high SNR, but their practical angular accuracy is still governed by aperture. Longer baselines give sharper bearings; the trade is that grating lobes appear for spacing greater than λ/2, requiring careful array design and potentially ambiguity resolution logic in the DF software. The angular uncertainty σ_θ from a single bearing measurement translates to a lateral position error at the emitter of approximately:

σ_lateral ≈ R · σ_θ      (σ_θ in radians, R = emitter range)

Example: σ_θ = 2° = 0.035 rad, R = 50 km
  σ_lateral ≈ 50,000 × 0.035 = 1,750 m

→ AOA fixes from a single 2° DF sensor at 50 km range have
  lateral uncertainty of ~1.75 km before geometry effects.

This range-scaling is the fundamental reason that AOA-only geolocation accuracy degrades rapidly with emitter distance, while TDOA (which measures range difference, not angle) degrades more slowly. For direction finding network architecture, sensor deployment close to the expected emitter area is the primary lever for improving AOA-based CEP.

Geometric dilution of precision (GDOP) for SIGINT networks

GDOP is the multiplier by which the measurement-domain error is amplified into the position domain. The same 10 m of TDOA range-difference uncertainty can produce a 15 m position error under excellent geometry or a 200 m position error under poor geometry. Understanding and controlling GDOP is the most powerful lever available to the SIGINT planner, because it does not require improving hardware, signals, or synchronization — only sensor placement.

The formal definition starts from the measurement equation linearized about the current position estimate. Let h(x) be the vector of predicted measurements (TDOA range-differences, AOA bearings) as a function of emitter position x. The Jacobian H = ∂h/∂x maps a small position change to the corresponding measurement change. Given a measurement covariance R, the position covariance from weighted least squares is:

P_x  =  (H^T · R⁻¹ · H)⁻¹

GDOP  =  √( tr(P_x) / σ_meas² )

Where σ_meas² is a reference measurement variance (e.g., c² · σ_τ²
for a TDOA system). GDOP is dimensionless; it is the factor by which
measurement noise propagates into position noise.

Geometrically, GDOP is lowest when the constraint surfaces (TDOA hyperbolas, AOA bearing rays) intersect at right angles near the emitter. When two TDOA hyperboloids are nearly tangent — which happens when the emitter is nearly equidistant from two sensor baselines — they carry little independent information and their intersection is poorly defined. Collinear geometry is the worst case: an emitter on the extension of the line connecting two sensors has one TDOA pair that provides no range resolution whatsoever in the along-baseline direction.

Practical GDOP values for a three-sensor TDOA network in 2D range from about 1.2 (excellent, sensors evenly distributed around the target) to 10+ (poor, sensors clustered or nearly collinear with the target). As a planning target, GDOP below 3 over the area of interest is desirable; GDOP above 5 in any part of the target area should trigger sensor repositioning or additional sensor deployment.

GDOP can be mapped over a geographic grid for a given sensor configuration, producing an accuracy footprint analogous to the GNSS PDOP maps used in precision-approach navigation. In SIGINT mission planning tools, this map lets the operator visually identify dead zones and hot spots before committing to a sensor deployment. The passive geolocation TDOA FDOA architecture discussion covers how these accuracy maps feed the collection-management workflow.

Sensor placement optimization for minimum CEP

Given a target area, a set of candidate sensor locations (constrained by terrain, road access, communication-link budget, and operational security), and a measurement error budget, sensor placement optimization finds the subset of candidate sites that minimizes CEP over the target area. This is a combinatorial optimization problem that is generally too large for exhaustive search when the candidate site set is large.

The objective function is usually one of:

• Minimize maximum GDOP over the target area (minimax): ensures no point in the area has catastrophically bad coverage.
• Minimize mean GDOP over the target area: optimizes average coverage at the cost of potentially bad corners.
• Maximize area within a GDOP threshold: ensures a specified fraction of the target area meets the accuracy requirement.

Gradient descent is straightforward to apply when sensor positions are treated as continuous variables: compute the gradient of the GDOP objective with respect to each sensor's coordinates and step in the descent direction. Because GDOP is non-convex in sensor positions, gradient descent finds a local minimum; multiple random restarts improve the chance of finding a global optimum. When sensor positions are restricted to a discrete set of candidate sites, the problem is combinatorial and gradient descent does not directly apply — but simulated annealing or genetic algorithms handle the discrete case well at moderate problem sizes.

Practical constraints shape what optimization can achieve. Terrain masking — high ground between a sensor and the emitter — creates a line-of-sight shadow that eliminates that sensor's contribution from certain directions; any site with severe masking toward the target area should be penalized heavily in the objective. Communication links from sensor to fusion center must have sufficient margin to carry the data rate required; a site with optimal geometry but no viable link is operationally useless. Road access for emplacement and logistics, power availability, and vulnerability to detection by the emitter's own sensing all further constrain the feasible set. The RF geolocation defense overview discusses how these constraints feed into system-level planning for tactical SIGINT deployments.

Multi-technique fusion: combining TDOA and AOA

TDOA and AOA carry geometrically orthogonal information. TDOA constrains the emitter along the direction perpendicular to the sensor baseline — it has high sensitivity in the cross-baseline direction and low sensitivity along the baseline. AOA constrains along the bearing ray — narrow in the angular direction, broad in range. When the two are fused, the combined constraint is tighter in both dimensions than either alone, producing a smaller error ellipse and a lower CEP.

The standard fusion approach for a static emitter is weighted least squares (WLS). Both measurement types are linearized about the current position estimate to produce a combined Jacobian:

H_combined  =  [ H_TDOA  ]     (TDOA rows: ∂Δr/∂x for each sensor pair)
               [ H_AOA   ]     (AOA rows:  ∂θ/∂x  for each DF sensor)

R_combined  =  blockdiag( R_TDOA,  R_AOA )

P_x  =  (H_combined^T · R_combined⁻¹ · H_combined)⁻¹

CEP50  ≈  1.18 · √(0.5 · (P_xx + P_yy - √((P_xx-P_yy)²+4P_xy²)))

The key insight from the covariance formula is that a well-oriented AOA measurement reduces the eigenvalue of P_x in the range direction — exactly the direction TDOA contributes least — while TDOA reduces it in the cross-baseline direction. The result is a more circular error distribution, which is both smaller in CEP and more predictable across the target area.

For a time-varying emitter — one that moves between measurements — a Kalman filter replaces the static WLS. The filter propagates the state estimate (position and velocity) forward in time using a kinematic model (constant velocity, constant acceleration, or a maneuver model with process noise), then updates the state when a measurement arrives from any sensor. TDOA and AOA measurements arrive asynchronously and are processed as individual updates. The Kalman filter handles the measurement timing naturally: the predict step extrapolates the position to the measurement time, and the update step applies the appropriate measurement Jacobian for the current estimated emitter position. Empirically, fusing TDOA and AOA reduces CEP50 by 30–60% relative to the better single-technique estimate, with the improvement being largest when the two techniques have comparable individual accuracies and their constraint surfaces intersect at large angles.

Communicating geolocation accuracy to intelligence consumers

A geolocation fix that reaches an analyst or decision-maker without its associated uncertainty is operationally dangerous. It invites the consumer to treat an imprecise estimate as a precise one, with consequences that range from wasted collection effort to targeting errors. Conversely, a fix accompanied by a well-characterized uncertainty region enables the consumer to make a calibrated decision: whether to task additional sensors, whether the target is close enough to a boundary to warrant further investigation, or whether the fix is good enough to act on immediately.

The minimum content of a geolocation report for intelligence use:

• Fix coordinate — WGS-84 latitude/longitude or MGRS grid reference, to the precision warranted by the CEP (reporting to 6 decimal places when the CEP is 2 km is false precision and should be avoided).
• Error ellipse — semi-major axis, semi-minor axis, and orientation (degrees clockwise from north). When map display is available, draw the ellipse rather than just listing its parameters; spatial reasoning about uncertainty is far easier visually than numerically.
• Confidence level — whether the ellipse is at CEP50, CEP90, or CEP95. Different organizations default to different levels; state it explicitly.
• Time of fix — the epoch at which the measurement was taken, in UTC. A fix that is 30 minutes old has very different operational value for a mobile emitter than for a static installation.
• Time validity — how long the fix is expected to remain accurate enough for its intended use. Stationary infrastructure: hours to days. Vehicle-mounted emitter: minutes to tens of minutes. Airborne emitter: seconds to minutes. This figure should be driven by expected emitter mobility, not by collection cadence.
• Technique and sensor count — whether the fix is from TDOA, AOA, or fusion; how many sensors contributed. A three-sensor TDOA fix from well-distributed collectors is more credible than a two-bearing AOA intersection at near-180° geometry.
• Geometry quality flag — a GDOP or condition-number indicator that tells the consumer whether the geometry was favorable or marginal. A GDOP flag of "poor" should trigger consumer skepticism about the stated CEP, since the error budget may not have fully accounted for all geometry effects.

Reporting standards for geolocation in intelligence products vary by organization and coalition. The core principle across all of them is the same: state what you know, state how confident you are, state how old it is, and state how that confidence degrades with time. Omitting any of these components forces the consumer to make an undocumented assumption — and undocumented assumptions are where operational errors originate.

On the common operating picture, error ellipses should be displayed at a consistent confidence level across all tracks, with a legend identifying that level. Mixing CEP50 ellipses from one source with CEP90 ellipses from another on the same display is a common integration failure that makes the picture look more precise than it is for some emitters and less precise than it is for others. When building or integrating a SIGINT system into a COP, standardize on CEP90 for displayed ellipses — it is the figure that corresponds most directly to worst-case planning, and it is what the consumer actually needs.