Combining messengers to constrain a source

Combine physical constraints rather than unlike numbers

A strain signal can constrain an orbital waveform model. Electromagnetic spectra can probe radiating material, while neutrino evidence can test particle-production and propagation hypotheses. Combining messengers is useful when the observations constrain different parts of a proposed physical explanation.

The quantities should first be connected by a model. Averaging a strain amplitude with a photon energy has no useful physical meaning. Likewise, multiplying two reported significance numbers does not by itself construct a joint source likelihood. Define the observables, their units, detector responses and shared parameters before combining them.

A host association or a distance estimate can be valuable new information, but it also carries its own selection and uncertainty. Two analyses using the same host or calibration reference may share those uncertainties. More data can improve an inference while still leaving an important parameter unidentifiable.

We will isolate that problem using a completely supplied timing example. Its values are invented exact mean delays in a static flat-frame model, not observed events, a cosmological timing formula or a measurement of a fundamental speed difference.

An arrival difference contains two contributions

For event i, let messenger 1 and messenger 2 travel the same known path length D_i at constant positive speeds v1 and v2 in the stipulated frame. Let τ_i be the emission time of messenger 2 minus that of messenger 1, in seconds. The difference between their calibrated arrival times is Δt_i = τ_i + D_i(1/v2 − 1/v1).

Choose a fixed positive reference length D0 and write D_i = d_iD0, with d_i dimensionless. Define κ = D0(1/v2 − 1/v1), which has units of seconds. The arrival model becomes Δt_i = τ_i + κd_i. A positive κ means messenger 2 takes longer to cross the same positive path.

The emission and propagation terms can therefore reinforce or partly cancel each other. Even a perfectly measured arrival delay need not be a direct measurement of κ when τ_i is unknown. Real clock, feature-definition, distance and response uncertainties would add further complications; they are held fixed here to reveal the structural issue.

This model deliberately excludes expanding spacetime, energy-dependent speeds, different paths and varying speeds along the path. Those require their appropriate propagation integrals. The simple D/v expression should not be transported into those settings without revising the model.

Two equations identify the parameters only in the restricted model

Supply two pairs: d1 = 1 with Δt1 = 3 s, and d2 = 3 with Δt2 = 5 s. Now impose the additional assumption that both events have one common intrinsic emission lag τ1 = τ2 = τ. This is a modelling restriction, not information contained in the supplied arrivals.

The two equations are 3 s = τ + κ and 5 s = τ + 3κ. Subtracting gives 2 s = 2κ, so κ = 1 s. Substituting back gives τ = 2 s. These parameter values reproduce both supplied delays exactly.

Under that same common-lag, common-propagation model, a new event with d = 2 would have predicted delay τ + 2κ = 4 s. That number is a conditional held-out prediction. No third event or delay measurement has been supplied.

The algebra does not establish that astrophysical sources share the assumed τ. Different engines, viewing directions, emission regions or transfer conditions can change emission timing. The scientific question is whether independent physical evidence and additional observations justify the shared-lag approximation for the selected events.

An equal-speed alternative fits the same supplied data

Release the common-lag assumption and choose κ = 0. This corresponds to equal propagation speeds in the supplied same-path model. Let τ1 = 3 s and τ2 = 5 s. The predicted arrival differences are again 3 s and 5 s, fitting both supplied pairs exactly.

The original common-lag solution and this equal-speed alternative therefore cannot be distinguished by these two supplied arrival values alone. One explanation attributes part of the difference to distance-dependent travel; the other attributes it to differing emission lags.

More generally, changing κ to κ + a and changing every free τ_i to τ_i − ad_i leaves each Δt_i unchanged, for a in seconds within the allowed physical parameter domain. This is an explicit direction of non-identifiability: the model predictions do not distinguish those parameter changes.

Adding many exact delays does not remove that ambiguity if a separate unrestricted emission lag is added for every new event. A prior or a population assumption can regularize the inference, but its influence should be reported. Mathematical precision under a chosen restriction is different from evidence that the restriction describes the sources.

A useful joint model makes its conditions testable

Additional observations can constrain the emission stage, test proposed feature correspondences or reveal source diversity. A physical bound on an intrinsic lag, a justified population distribution or a genuinely predictive emission model may add information that the arrival values alone lack. Each has to be motivated and its uncertainty retained.

The dated 2017 GW170817 and GRB170817A analysis explicitly used arrival timing together with distance and possible emission-time differences when constraining propagation. Its conclusions were conditional on those physical choices. The lesson here is that the assumptions belong with the bound; no numerical bound or current experimental status is reproduced.

A separate neutrino nondetection can test a proposed source spectrum and beaming pattern under its search response, rather than merely shrinking an unrelated timing parameter by declaration. The ANTARES, IceCube and Pierre Auger joint search illustrates why those energy, time and direction conditions matter.

When real observations have uncertainty, distinguish statistical variation, calibration uncertainty, common reference effects and model discrepancy. Correlations can change the joint uncertainty, while an unidentified parameter remains a structural problem even if random measurement errors become very small.

Report what was measured, which model linked the observations, which quantities were identified and what alternatives remain. The next lesson develops the last point further: a missing selected signal can constrain a model, but an uncovered direction can prevent a supposed all-direction limit.

Worked examples

Try the reasoning

1. Why does κ = D0(1/v2 − 1/v1) have units of seconds?

   Show the answer and reasoning

   Length multiplied by inverse speed gives time.

   Each inverse speed has units s m⁻¹ and D0 is in m; d_i is dimensionless.

2. What delay does the fitted common-lag model predict at d = 2, and is that a third observation?

   Show the answer and reasoning

   It predicts 4 s; this is not an observation.

   Use τ + κd = 2 s + 1 s × 2 under the same model restriction.

3. Do arbitrarily many exact arrival delays necessarily determine κ if each event has its own unrestricted τ_i?

   Show the answer and reasoning

   No.

   The change κ → κ + a, τ_i → τ_i − ad_i preserves every predicted delay within the allowed parameter domain.

4. Can two messenger analyses using the same uncertain host distance automatically treat that distance uncertainty as independent?

   Show the answer and reasoning

   No.

   The shared reference belongs in their joint uncertainty model; counting it twice as independent can misstate the constraint.

What to carry forward

• Combine messengers through defined observables, response models and shared physical parameters.
• An arrival delay contains emission and propagation terms whose separation needs identifying information.
• State and test the extra assumptions that turn a precise conditional solution into a physical constraint.