Integrity & RAIM
Not just being right — being able to prove it, and knowing the instant you can't
Why it matters
Accuracy without integrity is dangerous: a confidently wrong fix is indistinguishable from a correct one. Aviation, maritime, rail, and autonomous systems need a guarantee that errors beyond a safe limit will be caught in time. RAIM is the receiver-only form of that guarantee, and the conceptual root of SBAS, ARAIM, and all modern integrity architectures.
Prerequisites
The question accuracy can't answer
Everything in this course so far has chased one goal: get the fix close to the truth. But put yourself in the cockpit of an aircraft on an instrument approach, breaking through cloud toward a runway you cannot yet see, trusting GNSS to tell you where you are. Now "close" is not the question. The question is: if this fix were dangerously wrong right now, would I know?
That is a completely different property, and it has a name: integrity. A system has integrity not when it is accurate, but when it either keeps its error within a stated bound or tells you, in time, that it can't. The nightmare is not a large error. The nightmare is a large error that looks exactly like a small one — a fix sitting confidently fifty metres off the runway centreline with no hint that anything is wrong. Accuracy cannot save you here, because a single fix, examined on its own, has no way to know it is lying.
So how can a receiver possibly bound an error it cannot see?
The idea, before any mathematics
The escape is almost philosophical: you cannot check a claim against itself, but you can check it against other claims. One satellite gives you one measurement, and you have no way to know if it is honest. But ask more satellites than you strictly need, and their stories must be consistent — they are all, after all, describing the same receiver at the same instant. If one of them lies, it will disagree with the consensus of the others, and that disagreement is something you can see without ever knowing the truth.
This is the entire foundation of RAIM, and it rests on one word: redundancy. To fix a 2D position and a clock you need three satellites, with nothing left over to cross-examine. Add a fourth and suddenly the four must agree about three unknowns — there is one spare opinion, and any inconsistency surfaces as a leftover the solution cannot explain away: the residual. Add a fifth, and you have two spares — now not only can you tell that someone is lying, you can tell who, and throw them out.
Integrity does not come from better measurements. It comes from more of them than you need, so they can cross-examine each other. One spare lets you detect a liar; two let you exclude one.
But there is a subtle trap, and it is pure geometry. Some faults are easy to catch because they throw the measurements into obvious disagreement. Others — on a satellite sitting in a weak direction — can shove your position sideways while barely disturbing the consensus at all. Such a fault is nearly invisible to the residual yet damaging to the answer. The ratio of "how much position error" to "how much visible disagreement" is a satellite's slope, and the worst slope in your sky decides how large an undetectable error could be hiding. Bound that, and you have a protection level — a number you can honestly promise the pilot the error will not exceed.
Seeing it
Picture satellites ringed around you. Spoil one — nudge its range a little long — and watch the least-squares solution strain: it cannot satisfy the bad satellite and the honest majority at once, so it leaves a residual, and the residual grows with the bias. Cross a line and the alarm trips. Now picture the satellites bunched into one corner of the sky instead of spread around you: the same bias now slides your position a long way while the residual hardly stirs — the geometry has made the fault sneaky, and the protection level you must quote grows to cover it.
Try it yourself
Inject a bias onto any satellite and watch the two meters: the fault-detection statistic, and the protection level against the alert limit.
- Error (all-in)
- 50 cm
- Error (used)
- 50 cm
- Max slope
- 0.47
Accuracy asks how close the fix is; integrity asks whether you can trust it — and a confidently wrong fix looks identical to a good one until you have something to check it against. That something is redundancy: with more satellites than unknowns the measurements must agree, and their disagreement is the residual. Push the bias on a satellite and watch the top meter climb; cross the dashed threshold and the monitor declares a fault. With five or more satellites, turn on FDE: the monitor fingers the culprit, excludes it (marked with an ×), and the solution snaps back. The bottom meter is the real product of RAIM: the protection level, a geometry-bounded worst case for the error. As long as it stays under the alert limit, the operation is safe to fly. Now uncheck good geometry: clustered satellites have a huge slope, so a bias can shove the position far while barely moving the residual — the protection level balloons past the alert limit and integrity goes unavailable, even with no fault at all. That is the whole trade of safety-of-life GNSS: not just being right, but being able to prove you are.
Start clean: the detection meter sits low, the protection level sits under the alert limit, and the verdict is AVAILABLE. Drag the bias on a satellite upward and the detection statistic climbs until it trips the threshold — FAULT. Turn on FDE with five or more satellites and the monitor names the culprit, marks it with an ×, and recovers. Now the deeper lesson: uncheck good geometry so the satellites cluster. With no fault at all, the max slope leaps, the protection level balloons past the alert limit, and the verdict flips to UNAVAILABLE — the receiver refusing to certify a fix it cannot bound. Finally, try a small bias on a high-slope satellite that stays just under the detection threshold, and watch the actual error marker creep toward and past the protection level: that is precisely the hazardous case the whole machinery exists to make vanishingly unlikely.
Now the mathematics
RAIM lives on the same linearised least-squares system as single-point positioning. With pseudoranges and the geometry matrix (each row a line-of-sight unit vector plus a clock term), the position/clock solution and its post-fit residual vector are
where . The residual is what the solution could not absorb. Its size is the test statistic; a convenient form is
which, with no fault, is about the measurement noise (the UERE from the error budget). Declare a fault when exceeds a threshold , where is set from the chi-square distribution to hold the false-alarm rate down. The denominator is the redundancy: at it is zero, there is no residual, and no detection is possible — the mathematical statement of "you need a spare."
The protection level comes from the geometry. A bias on satellite pushes the horizontal position by and the test statistic by , where is the norm of the position row of the pseudo-inverse and is the residual sensitivity. Their ratio is the slope:
Read it plainly: the satellite with the worst slope is the one whose fault hides best in the residuals, so it sets the worst-case undetected error. Multiply that slope by the noise and a factor (which folds in the detection threshold and an allowance for missed detection), and you have the Horizontal Protection Level. The final decision is a single comparison:
If the protection level fits under the operation's Horizontal Alert Limit, you may use the fix; if not, the receiver must declare itself unavailable rather than offer an unguarded answer. Notice the slope and come from the very same that defines dilution of precision — integrity and geometry are two readings of one matrix.
How this appears in engineering
Classic RAIM is a receiver computing exactly this to flag a single faulty satellite for aviation's non-precision approaches, and it is visualised with the Stanford diagram, which plots the protection level against the true error over millions of samples and partitions the plane into nominal operation, system-unavailable, misleading, and hazardously misleading regions — the last of which must be driven below probabilities like per hour. Because standalone RAIM is limited (it assumes a single fault and needs good geometry), it grew up into richer architectures: SBAS (WAAS, EGNOS) broadcasts corrections and integrity bounds from a ground network; ARAIM uses multiple constellations and a published Integrity Support Message to handle multiple simultaneous faults for precision approach.
The same idea echoes everywhere once you see it. The Kalman filter's innovation test — rejecting a measurement that disagrees with the prediction by more than its covariance allows — is simply the recursive cousin of the RAIM residual check. Carrier-phase ambiguity validation in RTK is an integrity test on the integer solution. And in an era of spoofing, the residual that catches a lazy ephemeris fault is the same statistic that catches a counterfeit signal that forgot to stay consistent with its neighbours.
Check your understanding
- State the difference between accuracy and integrity in one sentence, then give an example of a system that is accurate but has no integrity.
- A receiver tracks exactly four satellites and one has a slow bias. What can RAIM do, what can it not do, and how many satellites would it need to do the thing it cannot?
- Two receivers quote the same 3 m accuracy, but one reports HPL = 25 m and the other HPL = 60 m against a 40 m alert limit. Which can fly the approach, and what physically differs between them?
- Explain why a fault on a high-slope satellite is the dangerous case, and what in the protection level formula is specifically there to defend against it.
Where to go next
You have now closed the loop from computing a position to trusting one. Integrity is the lens that reframes the whole course: DOP was always half of a protection level; the error budget was always the that scales the threshold; the Kalman innovation was always a fault test in disguise. From here the frontier is multi-constellation ARAIM and tightly-coupled integrity across GNSS and inertial sensors — because the systems we now trust with lives do not merely need to know where they are. They need to know when not to believe themselves.
Common misconceptions
- A more accurate receiver is automatically safer. Accuracy and integrity are different; a precise fix with an undetected fault and no protection level offers no safety guarantee.
- RAIM measures the actual error. It cannot see truth; it bounds the error statistically from geometry and residuals, and declares unavailable when it cannot guarantee the bound.
- More satellites only improve accuracy. Redundancy is what makes integrity possible at all — detection needs one spare measurement, exclusion needs two.
Key takeaways
- Integrity is a trustworthy error bound plus a timely alarm, not just a close fix.
- Redundancy makes measurements check each other; the residuals expose inconsistency.
- Fault detection needs n > 4 (one spare); fault exclusion needs n > 5 (two spares).
- A satellite's SLOPE is how much position error a hard-to-detect bias on it can cause.
- HPL below the Alert Limit means integrity is available; otherwise the operation is denied.