A shared reference removes some terms and creates shared uncertainty
Two receivers observe several satellites at the same epoch. Subtracting their measurements in a carefully defined order can remove receiver and satellite clock terms. That is a powerful simplification, but the resulting differences are not automatically independent or free of atmospheric, multipath and ambiguity problems. The reference satellite appears repeatedly in the error account.
This guide builds the double difference from four observations, checks a small clock example, and propagates the covariance created by a shared reference. It then changes the reference satellite without changing the underlying evidence. The purpose is to understand what relative GNSS processing does, not to turn a quality label into a promise of survey accuracy.
All clocks, residuals, wavelengths and uncertainty levels in the numerical cases are invented. The equations are restricted teaching models for compatible same-system observations. They do not configure a receiver, establish a legal position, or replace current signal and processing documentation.
Geodesy subject guide · Learning that transfers into life · How X Works hub
Choose your reading route
Route 1: Build the difference from observations
Route 2: Work clock cancellation and geometry
Route 3: Propagate shared-reference uncertainty
Route 4: Check reference changes and ambiguity
Route 5: Transfer and report
Full chapter index · Independent worked case · Questions and practical checks
Chapter index
Build the difference from observations
Label all four observations before subtracting
Write the clock terms in compatible units
Work clock cancellation and geometry
Check clock cancellation with small fictional values
Notice that geometric information remains
List what does not cancel exactly
Propagate shared-reference uncertainty
Calculate variance from the four original errors
Derive covariance from a shared reference satellite
Check a sum and a contrast of correlated differences
Check reference changes and ambiguity
Change the reference without inventing new information
Keep ambiguity continuity separate from clock cancellation
CHAPTER 1 OF 12
1. Label all four observations before subtracting
Let A and B identify two receivers and R and S identify two satellites. At a compatible epoch, there are four observations: yAᴿ, yBᴿ, yAˢ and yBˢ. Keep both indices visible. A double difference is not simply the difference between any two numbers labelled GNSS; it is a particular linear transformation involving receiver and satellite relationships.
Define the between-receiver single difference for satellite S as ΔS = yBˢ − yAˢ. Define ΔR similarly. Our double difference is DSR = ΔS − ΔR. Expanding gives yBˢ − yAˢ − yBᴿ + yAᴿ. Other sign conventions are valid if used consistently. Writing the expansion at the beginning prevents a silent switch of receiver order halfway through an uncertainty or ambiguity calculation.
The observations must have compatible time, units, signal definitions and processing conventions. A measurement at B several seconds later than A does not automatically share the same clock and geometric terms. Different frequencies or constellations can introduce bias relationships absent from the simple model. The labels are physical information, not decorative superscripts.
Start with four fictional observations on paper and colour the receiver and satellite terms separately. Then perform one subtraction at a time. This concrete approach helps learners see why some terms cancel and others remain. Only after the relationship is stable should the calculation add carrier ambiguities, atmosphere and a covariance matrix. More sophisticated notation is useful when it protects these meanings rather than concealing them.
CHAPTER 2 OF 12
2. Write the clock terms in compatible units
Use a simplified observation model yᵣˢ = ρᵣˢ + Cᵣ − Cˢ + other terms. The geometric range is ρ. Receiver clock contribution Cᵣ and satellite clock contribution Cˢ are already expressed as lengths, obtained by multiplying the relevant clock offsets by the speed of light. A time offset in seconds cannot be added directly to a range in metres without that conversion.
For a code model the ionospheric term has one sign; for a carrier-phase range model it has the opposite leading sign and an ambiguity contribution is present. The clock-cancellation logic is the same in this simplified representation. We first isolate it so that a learner can check the cancellation without confusing a clock term with an integer-cycle term.
For one satellite S, the between-receiver difference contains ρBˢ − ρAˢ + CB − CA. The satellite clock term cancels because the same Cˢ appears with the same sign in both compatible observations. The receiver-clock difference remains. Subtracting the corresponding single difference for satellite R removes that shared receiver-clock difference.
Exact cancellation here belongs to the stated common-term model. Actual processing must handle measurement timing, signal and hardware conventions appropriately. ESA Navipedia’s RTK explanation develops double-differenced carrier observables and the role of ambiguities. We use its basic observation structure without adopting historical example performance ranges as universal current guarantees. A derivation of clock cancellation is not yet a complete positioning quality assessment.
CHAPTER 3 OF 12
3. Check clock cancellation with small fictional values
Let the receiver clock contributions be CA = 10 metres and CB = 14 metres. Let satellite clock contributions be CR = 3 metres and CS = 8 metres. Ignore all other terms temporarily and examine only Cᵣ − Cˢ. The four clock contributions are A–R: 7, B–R: 11, A–S: 2 and B–S: 6 metres.
The receiver difference for R is 11 − 7 = 4 metres. For S it is 6 − 2 = 4 metres. Their double difference is 4 − 4 = 0. We did not measure the clocks as zero. We constructed a quantity insensitive to these particular common clock contributions. A transformed observable can remove a nuisance parameter without determining its original value.
Reverse the receiver order in both single differences and the resulting double difference changes sign consistently. Reverse it in only one single difference and the clock result becomes 8 or −8 metres rather than zero under the chosen ordering. The failure is a bookkeeping error, not evidence of unexpected clock physics. Expand the four terms to locate the first incorrect sign.
Now let the receiver contribution at B change between the epochs used for R and S. The same 4-metre receiver difference no longer appears twice, so the simple cancellation does not hold. This changed-context test exposes why timing belongs in the observation contract. Real methods can model asynchronous effects, but their successful treatment requires more than copying the simultaneous-observation formula.
CHAPTER 4 OF 12
4. Notice that geometric information remains
The geometric part of DSR is (ρBˢ − ρAˢ) − (ρBᴿ − ρAᴿ). It generally does not vanish, because the two satellites are viewed along different directions. That remaining dependence on the receiver separation is precisely what makes the relative-position problem useful. If every relevant term vanished, the double difference would contain no position information.
For a local linear illustration, suppose the range difference to S changes by kS times a small baseline displacement d, while the range difference to R changes by kR times d. The double difference changes by (kS−kR)d. Take fictional sensitivities kS = 0.6 and kR = −0.2. A 10-millimetre displacement along the selected direction produces an 8-millimetre double-difference change in this one-dimensional model.
If the sensitivities were nearly equal, the double difference would respond weakly to that displacement direction. More repeated observations at the same weak geometry do not necessarily resolve the missing component well. Actual three-dimensional positioning uses the appropriate geometry and several observations, with additional unknowns and constraints. One double difference is not a complete baseline solution.
The existing GPS range-and-clock guide explains how satellite geometry affects a position estimate. Use it to connect this observation-level sensitivity with a larger estimation system. Keep the stages distinct: clock cancellation simplifies the model, geometry determines sensitivity, ambiguity handling affects carrier interpretation, and an independent reference assesses the resulting position claim. None of those jobs can be replaced by the others.
CHAPTER 5 OF 12
5. List what does not cancel exactly
Nearby receivers may experience similar atmospheric contributions, so differencing can reduce part of those effects. Similar does not mean identical. The rays travel through different paths, and the degree of common behaviour depends on geometry, conditions and the observation model. A fixed maximum baseline length cannot be inferred from the clock-cancellation algebra alone.
Multipath is especially tied to each antenna’s surroundings. A reflection beside A need not be repeated beside B, even when they observe the same satellite. Differencing can carry that local contribution into the relative observable rather than eliminate it. Receiver noise also remains and combines according to the signed weights. A small undifferenced noise does not automatically become smaller after subtraction.
Antenna reference points, radome identities and calibration conventions still matter. So does the reference coordinate used to locate the base station. A precise relative baseline added to an incorrect base position can yield a correspondingly displaced absolute position. The baseline estimate and its global placement are separate parts of the evidence chain.
Signal-dependent hardware biases and mixed-system combinations require their own models. The restricted same-system, compatible-signal explanation here should not be stretched into a statement that every instrumental bias cancels for arbitrary receiver pairs. Make a surviving-term ledger before interpreting residuals. A term may be reduced, modelled, estimated or left uncertain; those are different descriptions, and a reliable report says which actually occurred.
CHAPTER 6 OF 12
6. Calculate variance from the four original errors
Assume, for this fictional example only, that each of the four original observation errors is independent with standard uncertainty σ. The double difference has coefficients +1, −1, −1 and +1. Its variance is therefore 4σ² and its standard uncertainty is 2σ. The negative signs do not cancel independent noise, because variance uses squared coefficients.
Let σ = 2 millimetres for each original carrier-range observation. Each between-receiver single difference has variance 2×2² = 8 square millimetres. A double difference built from two independent single differences has variance 8 + 8 = 16 square millimetres, giving standard uncertainty 4 millimetres. This is an observation uncertainty, not a three-dimensional coordinate accuracy.
If the original observations are correlated or have unequal uncertainties, the simple factor of two no longer captures the full result. Use the complete covariance propagation for the actual linear transformation. Shared receiver processing, signal characteristics or atmospheric modelling may change the relationships. Their presence does not establish a particular numerical covariance without evidence.
This calculation is a useful check on a common misconception: differencing removes everything common, so it must reduce every error. It removes specific exactly common terms while adding independent error contributions. A processing strategy can be valuable because it eliminates troublesome nuisance parameters even when the transformed measurement noise increases. Evaluate the full estimation problem rather than judging a technique from only one error term.
CHAPTER 7 OF 12
7. Derive covariance from a shared reference satellite
Add a third satellite T. Form DSR = ΔS − ΔR and DTR = ΔT − ΔR. Even if the three single differences ΔR, ΔS and ΔT have independent errors, the two double differences share the same ΔR error. When that reference error shifts, both double differences shift together. The outputs are correlated because their construction reuses an input.
With independent original errors of standard uncertainty 2 millimetres, each single-difference variance is 8 square millimetres. Both double-difference variances are 16, while their covariance is Var(ΔR) = 8 square millimetres. The correlation is 8/√(16×16) = 0.5. This correlation follows from the stipulated equal independent-input model, not from a universal empirical law for all GNSS data.
The covariance matrix is therefore rows [16, 8] and [8, 16], in square millimetres. Keeping only its diagonal would incorrectly describe the two transformed observations as independent. A later least-squares baseline estimate should account for the shared-reference structure or use an equivalent formulation that preserves the information in the original observation model.
The same principle applies whenever many reported quantities subtract the same uncertain reference. Different column names do not make them independent. Trace each output back to its input weights. That exercise often reveals dependence more clearly than looking for correlation in a short residual scatter plot, especially when the shared contribution has been partly absorbed by other fitted parameters.
CHAPTER 8 OF 12
8. Check a sum and a contrast of correlated differences
For the two fictional double differences, consider their mean M = (DSR + DTR)/2. Its variance is (16 + 16 + 2×8)/4 = 12 square millimetres. The standard uncertainty is √12, about 3.464 millimetres. Treating the two double differences as independent would give variance 8 and standard uncertainty about 2.828, understating this mean’s uncertainty under the stipulated model.
Now consider their contrast Q = DSR − DTR. Its variance is 16 + 16 − 2×8 = 16 square millimetres, giving standard uncertainty 4 millimetres. Here the common reference error cancels. Ignoring covariance would instead give variance 32 and overstate this contrast’s uncertainty. The same positive covariance affects a sum and a difference in opposite directions.
Neither result says the data became inherently better or worse in every sense. The target quantity changed. A shared reference is relevant to the mean but absent from the contrast. This is why one generic uncertainty number should not be pasted beside every derived combination. Propagate the transformation appropriate to the actual decision.
For a learner, predict which reference term survives before touching the numbers. Expand M and Q in terms of ΔR, ΔS and ΔT. The mean contains −ΔR once, while the contrast is ΔS−ΔT and contains no ΔR. That physical bookkeeping provides an independent check of the covariance arithmetic and makes a matrix formula easier to understand.
CHAPTER 9 OF 12
9. Change the reference without inventing new information
Suppose the reference satellite changes from R to T. The new differences can be written DRT = −DTR and DST = DSR − DTR. They are linear transformations of the original pair. If all relevant observations and parameters are transformed consistently, changing the reference basis does not create a new independent dataset or a new physical displacement.
Write the old vector as [DSR, DTR]ᵀ. The new vector [DRT, DST]ᵀ is obtained with matrix rows [0, −1] and [1, −1]. Its covariance must become T C Tᵀ, using T for this transformation matrix. In the equal-variance fictional example, that calculation gives the same numerical matrix [16,8;8,16], though the observation meanings have changed.
That numerical equality depends on the symmetry of the example. With unequal input uncertainties or correlations, the new covariance entries need not match the old ones. Copying the original diagonal error bars after changing reference is not a general rule. Transform ambiguity parameters and other relevant quantities consistently as well as the observation values.
A real processing sequence may also change satellite availability, weights, ambiguity arcs or rejected observations when its reference changes. Those changes can alter the information actually used. Distinguish a pure algebraic re-expression of the same evidence from a different solution configuration. A jump coinciding with a reference switch requires inspection of that full change, not an automatic conclusion that the ground moved or that the reference choice alone explains everything.
CHAPTER 10 OF 12
10. Keep ambiguity continuity separate from clock cancellation
For compatible carrier-phase range observations, the simplified model includes λN, where N is an ambiguity parameter and λ is the signal wavelength. Double differencing combines the ambiguity terms; it does not automatically determine them. Under the appropriate same-frequency and bias assumptions, the resulting double-difference ambiguity has an integer structure useful for relative positioning.
A float ambiguity estimate is not the same thing as a validated integer solution. The existing RTK terminology guide explains fixed and float status without treating fixed as guaranteed truth. The carrier-continuity guide explains how a loss of tracking can change the ambiguity arc. Those issues remain relevant after receiver and satellite clocks have been removed from the simplified equation.
Changing the reference satellite also changes the representation of the ambiguity vector. A correctly transformed parameter set can preserve the same physical model. Resetting or mishandling an ambiguity arc instead changes the estimation problem. Keep a record of reference changes and tracking interruptions so that a discontinuity can be traced to its observation history.
Do not extend this simple integer discussion to arbitrary mixed-frequency, mixed-constellation or uncalibrated receiver combinations. Their bias conventions can require additional treatment. The first-principles question is which common terms the differencing actually removed and which remain in the parameterisation. A mathematical cancellation is trustworthy only when its physical assumptions match the observations being combined.
CHAPTER 11 OF 12
11. Distinguish algebraic closure from independent validation
With three single differences, DSR − DTR = DST by construction. Rearranging gives a zero closure expression. If all three double differences were calculated from the same underlying values, that closure is an identity. It can catch a transcription or inconsistent-sign error, but it is not independent evidence that the original observations are unbiased.
For fictional single differences ΔR = 1, ΔS = 5 and ΔT = 8 millimetres, DSR = 4, DTR = 7 and DST = −3. Then 4 − 7 − (−3) = 0. Add a 6-millimetre error to ΔR: the first two double differences become −2 and 1, while DST remains −3. Closure is still zero. The shared reference error did not break the identity.
This transfer check is valuable because a closed loop looks persuasive. The check’s independence comes from how the data were obtained, not from the visual shape of a triangle or the word closure. Survey levelling loops with independently observed segments have a different evidential structure from three algebraically dependent quantities created from the same inputs.
Use the dedicated levelling-loop guide when examining a physical survey network, and preserve its observation assumptions. For GNSS, independent baseline checks, suitable control and defensible alternative processing can address different weaknesses. Do not count a derived redundant combination as another independent measurement simply because it occupies a third row in a table.
CHAPTER 12 OF 12
12. Write a relative-position claim with its reference intact
A complete account identifies both receivers, observation epochs, satellite and signal identities, units, reference choice and processing version. It states the differencing order and the treatment of clocks, atmosphere, antenna models and carrier ambiguities. If the reference changes, the report explains whether the same information was transformed or the underlying solution configuration also changed.
Attach the correct covariance to the transformed observations and any derived baseline. State the source and reference of the base coordinate before turning a relative result into an absolute position. Report independent checks separately from algebraic identities and training-fit residuals. A small internal residual can coexist with a common reference error or an incorrect global placement.
For the fictional equal-noise case, four independent 2-millimetre observation uncertainties produce a 4-millimetre double-difference uncertainty. Two differences sharing a reference have covariance 8 square millimetres. Their mean and contrast therefore need different propagation. These conclusions are exact within the stipulated model and should not be advertised as receiver performance claims.
The gold standard is a relative measurement whose reference and dependence structure remain visible. Differencing is useful precisely because it changes the problem. Understanding that change lets a reader retain the clock-cancellation benefit while checking the atmospheric, geometric, ambiguity and reference-coordinate questions the transformation did not settle.
For a final evidence exercise, place the same incorrect base coordinate into two otherwise different baseline solutions. Their absolute rover positions may agree because both inherited the same placement error. That agreement does not independently validate the base. A reference check must address the reference itself, through information not already imposed on both solutions. Conversely, two solutions expressed in different frames may disagree numerically while describing the same physical point after an appropriate transformation. Identify the comparison model before interpreting the discrepancy.
A useful handover separates three products: the original observations and their identities, the relative solution with its covariance and assumptions, and the absolute placement using a named base coordinate and epoch. Keeping those products connected but distinct lets another reader investigate an observation problem without silently changing the frame, or update a reference coordinate without pretending that new receiver measurements were collected.
Independent worked case and evidence checks
A reserved fictional test has independent single-difference errors uR, uS and uT with standard uncertainty 3 millimetres each. Then each double difference has variance 18 square millimetres and the two sharing R have covariance 9. Their mean has variance (18+18+18)/4 = 13.5 square millimetres, while their contrast has variance 18. Explain the shared-reference cancellation before calculating either result.
Now assume the reference single difference has standard uncertainty 5 millimetres, while the other two remain at 3. Double-difference variances become 34 each and their covariance is 25. The mean variance is (34+34+50)/4 = 29.5, but the contrast variance remains 34+34−50 = 18. The contrast is unchanged because it contains no reference input at all.
Changed context: the original observations are correlated across receivers. Neither the equal-noise factor nor these simple sums are automatically valid. Begin with the appropriate original covariance and apply the actual differencing matrix. A correct answer identifies the assumption that changed instead of preserving a convenient formula simply because the number of satellites stayed the same.
Worked comparison
| Operation or quantity | Fictional result | What it checks |
|---|---|---|
| Clock-only double difference | 0 m | Specified common clock terms cancel |
| Four independent 2 mm input uncertainties | 4 mm output standard uncertainty | Independent noise does not cancel |
| Two differences sharing one equal-quality reference | 8 mm² covariance | Reference reuse creates dependence |
| Mean of those two double differences | √12 mm standard uncertainty | Shared reference remains |
| Contrast of those two double differences | 4 mm standard uncertainty | Shared reference cancels |
| Closure from the same three inputs | Exactly zero algebraically | Identity, not independent accuracy evidence |
Questions and practical checks
Does double differencing remove all GNSS errors?
No. It removes particular common terms under compatible observation assumptions. Geometry, atmospheric differences, multipath, noise, antenna effects and ambiguity questions remain.
Why are double differences correlated?
Several outputs reuse the same uncertain reference-satellite observation or single difference. That shared input creates covariance even when the original input errors were independent.
Can I choose a new reference and keep the old error bars?
Only if the full transformation and covariance justify it. A pure reference change requires consistent observation, parameter and covariance transformation; unequal input quality generally changes the entries.
Does a zero closure prove the observations are correct?
Not when closure follows algebraically from the same original inputs. It can check consistency of calculation but will remain zero even with some shared errors. Independent measurements answer a different question.
Is a precise baseline enough for an absolute position?
No. The base coordinate, reference frame, epoch and physical point must also be correct and appropriately uncertain. Relative precision does not independently verify global placement.
Continue with the connected subjects
Establish the original observation meanings
GPS ranges, clock bias and satellite geometry
RINEX epochs, signals and preserved measurements
Follow the remaining reference and validation checks
Carrier continuity, ambiguity changes and displacement limits
Antenna reference points and phase-centre calibration
Base coordinates, reference frames and coordinate epochs
RTK fixed and float status without unsupported accuracy claims
Levelling-loop misclosure and independently observed survey segments
Sources and scope
ESA Navipedia: double-differenced carrier observations and ambiguity structure
ESA Navipedia: other GNSS observation combinations and their different purposes
IERS: coordinates, velocities and terrestrial reference realisation
The numerical cases in this guide are original fictional teaching examples, not measurements from a real site. This is an educational explanation, not a field safety instruction, engineering approval or an official certification. For a real project, use qualified practitioners, current product documentation and site-specific evidence. Sources were checked on 31 October 2026.
Connect the geodetic observation checks
These companions examine different transformations of the evidence. Follow the unresolved term: repeated propagation, long-term estimation, frequency dependence or shared references. Passing one check does not settle the others.
Multipath: repeated geometry and held-out correction checks. Test repeated signal-path structure without treating repeatability as independent accuracy evidence.
Station velocities: seasons, slope weights and temporal covariance. Distinguish a fitted rate from seasonal sampling and uncertainty that persists through time.
Ionosphere-free combinations: cancellation and surviving noise. Check what a frequency combination removes, preserves and amplifies.