GNSS Accuracy Decoded: Why Positioning is a Game of Probabilities

GNSS accuracy is not one fixed number; it is a statistical claim wrapped around receiver quality, satellite geometry, signal environment, correction state, antenna setup, coordinate frame, and test method.

Most sensors make accuracy feel simple. A temperature probe might specify a range, a resolution, and a maximum error. A GNSS receiver looks similar on a datasheet, but the number hides more assumptions. A receiver can output latitude and longitude with many decimal places while the real position is still meters away.

The reason is that satellite positioning is a ranging problem measured against something outside the device. Each fix depends on time measurements from multiple satellites, and each measurement carries noise, atmospheric delay, multipath, receiver uncertainty, and geometry effects. The result is not a single guaranteed point. It is a cloud of possible positions.

What Kind of Sensor Is GNSS?

The usual sensor checklist behaves differently for GNSS:

• Range: Global. A GNSS receiver can operate almost anywhere with sky view.
• Resolution: Extremely fine. Coordinates can be reported down to millimeter digits, which is a property of the output format rather than evidence of millimeter-level measurement.
• Accuracy: Variable. Depending on signal quality, correction state, antenna setup, and environment, the error can range from centimeters to dozens of meters.

Fine resolution alongside large real-world error already rules out a plus-or-minus reading. What sets GNSS apart from the other sensors on a robot or a vehicle is what it measures against:

GPS program documentation draws the same line from the other direction: the error of the signal in space is one quantity, and the accuracy a user obtains is another, shaped by satellite geometry, obstruction, atmospheric conditions and receiver quality.

So a GNSS accuracy specification carries a metric and its test conditions. It is not a hard error ceiling that holds in your deployment, which is why the rest of this guide is written in probabilities.

Accuracy vs Precision in GNSS

Accuracy and precision are often used as if they mean the same thing. In GNSS, mixing them up leads to bad receiver comparisons.

• Accuracy means closeness to the true coordinate. The question is: did the receiver land on the right place?
• Precision means repeatability. The question is: do repeated fixes cluster tightly together?

The Target Analogy

• High precision, low accuracy: All your shots hit the same spot, but that spot is far from the bullseye. In GNSS, this often points to a systematic error.
• Low precision, high accuracy: Your shots scatter around the target, but the average center is close to the bullseye. In GNSS, this usually points to random error.

A receiver can be precise but inaccurate. If the antenna offset is wrong, the base-station coordinate is wrong, or the map layer uses a different datum, every point may repeat neatly while staying shifted from the true location. A clean track overlay is therefore not proof of absolute accuracy; the coordinate-frame side of that problem sits in the accuracy chain below.

The reverse can also happen. A noisy receiver may scatter around the true coordinate. The average may be close, but each individual fix is too unstable for control, mapping, or repeatable measurement. For robotics, field mapping, and machine guidance, both properties matter: the receiver needs a stable output and the output needs to be tied to the correct coordinate frame.

CEP, RMS, and R95 Explained

Most confusion starts when two datasheets describe accuracy with different metrics. A "1 m" number can mean median performance, root-mean-square spread, or a 95% containment radius. Those are not interchangeable claims.

The practical comparison table is:

Under a simplified two-dimensional Gaussian assumption, these metrics convert with published multipliers. For the same receiver, a CEP number always looks smaller than a high-confidence containment number, so the smaller figure on a datasheet may simply be the lower-confidence one.

2DRMS sits outside that conversion table. It is a conservative shorthand, roughly twice DRMS, and it is not strictly equal to R95. Its actual coverage shifts once the error distribution stops being circular, so treat it as a containment convention rather than as a converted value.

The Conversion Cheat Sheet

Engineers often need to compare a datasheet using CEP against one using RMS or a 95% statement. Assuming a standard two-dimensional error distribution, these multipliers give a practical conversion starting point:

• Scenario: A module claims a single-point positioning accuracy of 1.2 m CEP.
• To DRMS: Multiply by 1.19 -> 1.43 m DRMS.
• To R95: Multiply by 2.08 -> 2.5 m R95.

Conversion factors after F. van Diggelen, “GNSS Accuracy: Lies, Damn Lies, and Statistics”, GPS World, Vol. 18, No. 1, 2007. The table assumes a Gaussian, circular horizontal error distribution. R67 and R95 are containment radii. DRMS is a root-mean-square measure whose coverage lands near 63–68% under the same assumption, so it sits close to R67 without being the same quantity.

Why Dice Explain GNSS Accuracy

Those multipliers work because many GNSS accuracy calculations treat the final error as approximately Gaussian. That assumption comes from adding many smaller error terms together.

The Dice Experiment

Imagine rolling one die. The result is uniform: 1 through 6 have equal probability. That is not Gaussian. Now roll three dice, record the sum, and repeat the experiment 100,000 times. The histogram forms a bell curve because there are many more ways to make middle values than edge values. The sum of three dice is still discrete and bounded, so it is not literally a Gaussian — but it already shows the shape that emerges once independent terms accumulate.

function random_die() {
    return floor(random(1, 6));
}

for (i = 0; i < 100000; i++) {
    s = random_die() + random_die() + random_die();
    histogram[s] += 1;
}

The Central Limit Theorem

The Central Limit Theorem explains why this matters. A GNSS position error is not one clean error source. It is the sum of satellite clock residuals, orbital uncertainty, ionospheric delay, tropospheric delay, signal effects, multipath, receiver noise, antenna effects, and local geometry.

Individual errors may not be Gaussian. Their sum often behaves close enough to a Gaussian distribution that CEP, DRMS, R95, and similar statistical tools become useful. The assumption is never perfect, but it is practical when the test environment and confidence metric are stated clearly.

Two limits come with that assumption. First, a histogram that looks Gaussian does not mean successive epochs are independent: GNSS errors are usually correlated in time, and Kalman filtering strengthens that correlation. A short, time-correlated log cannot stand in for long-term behavior.

Second, changing operating state mixes different error populations into one sample, and the combined sample no longer necessarily satisfies a single distribution model. That is the assumption behind the four conditions in the previous section, and it is why a stated fix mode does not substitute for them.

The GNSS Accuracy Chain

If GNSS accuracy is set by conditions rather than by the device alone, the next thing to pin down is which conditions, and in what order. The path from a raw measurement to a comparable accuracy figure has a main line, one branch, and a set of reference conditions that sit outside the solution entirely:

Every GNSS position passes through the main line. Only RTK and DGNSS operation passes through the branch. The reference conditions never enter the solution at all; they decide whether two figures are comparable in the first place.

An accuracy number is comparable only when it is published with its observing conditions, its solution path, and its reference conditions.

How to Read a GNSS Accuracy Claim

Before treating an accuracy line as a requirement, check the conditions behind it. Each row maps to a link in the chain above:

Metrics describe the shape of the result. The error budget explains where that shape came from: satellite clock, ephemeris, ionosphere, troposphere, multipath, receiver noise, and geometry.

Why DOP Changes Position Error

DOP stands for Dilution of Precision. It is not a separate error source like multipath or atmospheric delay. It is a geometry multiplier. When satellites are spread across the sky, the receiver can separate timing errors into a cleaner position solution. When satellites cluster in one part of the sky, the same ranging error expands into a larger position error.

The simplified relationship is:

DOP is a family, not a single number. Under a simplified covariance model, each member isolates one part of the geometry:

• HDOP describes the horizontal geometry effect on latitude and longitude.
• VDOP describes the vertical geometry effect on altitude.
• PDOP covers three-dimensional position, combining the two above.
• TDOP covers the receiver clock term.
• GDOP is the overall figure, combining position and time.

Under that model they are related, not independent:

GDOP is the single figure that summarizes how far the current satellite geometry degrades both position and time. A receiver reporting a comfortable HDOP can carry a much larger VDOP at the same instant, and GDOP reflects both plus the clock term, which is why a DOP value quoted without saying which one is not a usable number. For the underlying derivation, see the ESA GNSS Data Processing handbook.

The Rule of Thumb

This explains why satellite count alone is a weak quality indicator. Twelve satellites with poor geometry can perform worse than eight satellites spread across a better sky.

Vertical accuracy is usually weaker than horizontal accuracy because satellites are above the receiver, not below it. The geometry has less vertical leverage, so altitude error often grows faster than horizontal error under the same ranging conditions.

In open sky, HDOP may sit around 0.7. In an urban canyon or beside tall structures, HDOP can rise toward 4.0, and horizontal accuracy can degrade roughly in proportion even before multipath is considered. Near buildings, tree lines, slopes, machinery, or greenhouse structures, HDOP can climb while reflected signals increase at the same time. The receiver may still produce a coordinate, but the probability cloud is larger and less symmetric.

What RTK Changes About GNSS Accuracy

RTK does not make the probability problem disappear. It changes which error terms remain dominant.

With ordinary single-point GNSS, the rover depends on broadcast satellite data, atmospheric models, its antenna environment, and its receiver design. With RTK, a nearby reference receiver observes many of the same satellite-side and atmospheric errors. The rover uses correction data to remove much of the shared error and then solves carrier-phase ambiguity for a centimeter-level fixed solution. That changes which observations are usable and which error terms dominate what is left.

That is why an RTK receiver can report a much tighter horizontal result than a standalone receiver under the same sky. It is also why RTK is sensitive to correction age, baseline length, cycle slips, multipath, and antenna installation. A fixed solution is a strong state, not a guarantee. The RTK Trick explains how double differences and integer ambiguity resolution produce that state—and why a Fixed solution can still be wrong.

Fixed is a solution-state label, not an error measurement. Two receivers both reporting Fixed can produce different error distributions, so a position is worth publishing with its state attached: fix mode, correction age, covariance, timestamp and validity flag. Downstream systems can then weight the position by that state instead of trusting a nominal figure.

Downstream, a centimeter-level fix is a component rather than a result: robot control stacks still have to handle state changes and obstruction, and a mapped boundary still depends on the coordinate frame and the field workflow.

Testing GNSS Accuracy in the Field

Reading the datasheet is the first step. Testing the receiver in the environment where it will actually work is the second.

The Golden Rule of Testing

The reference uncertainty must be documented and materially smaller than the error being evaluated. For a meter-level receiver, a surveyed control point or a higher-grade RTK system will usually meet that bar. For centimeter-level validation, the reference coordinate, antenna mount, correction source, and logging method all need tighter control, and the reference uncertainty itself has to be stated rather than assumed.

A practical field test should log:

• Timestamp, latitude, longitude, altitude, and fix mode.
• Satellite count, HDOP, correction age, and RTK float/fixed state.
• Antenna type, antenna height, mount position, and surrounding obstructions.
• Reference coordinate, coordinate frame, and map layer or survey control source.
• Whether the test is static, dynamic, repeated-loop, or revisit-based.

No Budget for High-End Gear?

Without a survey control point, repeated tests still help: log a static point cloud, drive the same route several times, revisit the same target at different times of day. Those runs will not prove absolute accuracy, but they expose precision, drift, multipath sensitivity, and correction-state stability.

Once a test produces Fixed / Float states, RMS figures, CDF curves, and degraded route segments, reading those layers together becomes its own problem. For that evaluation layer, see RTK Accuracy Trap.

Conclusion

A GNSS accuracy figure is comparable only when the metric, the solution mode, the test conditions, and the reference are known. Without them, 1.2 m or 1 cm is marketing shorthand, not an engineering input.

Zero Jiang

Founder, Kalmix

Dedicated to making high-precision GNSS positioning accessible and reliable for global developers building autonomous systems and robust RTK hardware.