Selective Propagation

Version

v4 · 2026-08-30 · History

Prime #

1451

Origin domain

Physical Sciences

Subdomain

transport and filtering → Physical Sciences

Related primes

Propagation

Core Idea

Selective propagation occurs when a mixed population travels one shared path and the path decides, component by component, what carries on. Each component holds some property — a wavelength, a grain size, a variant's transmissibility, a claim's memorability — and that property fixes its chance of continuing through the next stretch of medium or the next hop of a chain. Survivors are not altered by the journey; the mixture is. What is found far along the path is a biased draw from what set out, and the bias accumulates with distance covered rather than with time elapsed. ^[1]

The identity is narrower than the ordinary phrase suggests. Five things must hold together: an input heterogeneous in some property, a common path rather than a set of separate routes, a rule mapping that property to probability or distance of continued passage, progressive travel that accumulates exposure, and an output whose composition measurably differs from the input's. The sorting has to happen during passage; a mixture divided once at the entrance and sent onward intact is a different pattern with different remedies.

What makes the abstraction worth naming is the epistemic consequence. Anyone stationed downstream observes survivors, and survivors carry two signatures superimposed: what the source put in, and what the channel let through. Treating the far-field mixture as though it were the source's own is the recurring mistake, and gathering more observations at the same station cannot repair it, because every additional observation arrives through the same filter.

Structural Signature

Heterogeneous input → one shared path → per-component survival rule → attrition accumulating over distance or hops → composition-shifted output, every survivor individually unchanged.

The weight of the formula sits on the third arrow. The rule belongs jointly to the medium and to the attribute it responds to, and until both halves are named the pattern has not actually been identified.

Recurring features:

• A read property. The population varies along some dimension the medium actually responds to — frequency, size, charge, reactivity, infectivity, phrasing — and components alike on that dimension fare alike however else they differ.
• One path, not many. Everything shares a medium, a channel or a chain of relays; this is not a dispatcher assigning separate destinations at launch.
• A survival rule, thresholded or graded. Passage may be all-or-nothing at a boundary or a continuous probability per unit of exposure, and the abstraction is indifferent between them.
• Multiplicative accumulation. Survival across a whole path is the product of survival across its parts, which makes the shift exponential in distance and geometric in hop count.
• Attrition rather than transformation. Components are absorbed, scattered, settled, blocked, forgotten or declined, and those that continue are not converted into other components along the way.
• Reach as the discriminator. Two arrivals may travel at identical speed and still constitute a shifted sample, because what separates them from the missing is how far they could get, not how quickly.
• A source no single station recovers. From one downstream measurement the input composition and the channel's response are confounded, and no amount of sampling at that station pulls them apart.

The last feature yields the entry test in its negative form: if nothing was preferentially lost, then whatever else the path did to the population, this is not the pattern.

What It Is Not

This is not an accusation. Describing a downstream population as composition-shifted is arithmetic about a mixture, not a complaint: nothing need have gone wrong, no one need have chosen the survivors, and the output is not thereby worse than the input. A sunset reddens by exactly this mechanism, and the sky is in perfect order. ^[2]

Nor does anything here require an agent. The medium need not know what it is doing, and in the paradigm cases there is nothing present that could know: glass absorbs, hillsides drop their coarse grains, immune systems clear some lineages sooner than others. Where a person or an institution does the deciding, that is one realisation among many rather than the definition.

Nor must the read property be one anyone cares about. The channel responds to what it responds to, and any link between the surviving attribute and importance, accuracy, quality or merit is contingent and has to be argued separately. The structure supplies no such link and should not be used to smuggle one in. ^[3]

Nor is severity the same thing as selectivity. A path that destroys ninety-nine percent of everything at equal rates is not an instance; a path that removes four percent of one component and three percent of another is, weakly. The differential constitutes the pattern, not the magnitude, which is why some very lossy channels can be read almost at face value and some very gentle ones cannot.

Nor is the claim that survivors are unrepresentative in every respect. They are unrepresentative in the read property and in whatever correlates with it, and may be an excellent sample of everything else, which is precisely what makes partial correction feasible.

Finally, none of this is confined to loss. Where some components are relayed, boosted or re-emitted more readily than others, the same structure holds with passage probabilities above rather than below average; what matters is that continued passage is differential, not the sign of the difference.

Broad Use

Optics and radiation. Atmospheric scattering strips short wavelengths from a direct sunbeam far faster than long ones, so the transmitted beam reddens as path length grows while every photon in it moves at the same speed. A polychromatic X-ray beam hardens on its way through matter as low-energy photons are absorbed preferentially, so mean beam energy rises with depth. In a multimode fibre, higher-order modes lose power faster than low-order ones, and the far end of a long link carries a mode distribution the launch never had. ^[2]

Earth and environment. Anelastic attenuation removes high frequencies from a seismic wavetrain, so a distant station records a smoother source than a near one and one earthquake looks like two different events at two different ranges. Airborne ash and river-borne sediment both fine with distance as settling takes the coarse fraction first. Within an atmospheric plume the more reactive species are consumed en route, so the downwind mixture drifts toward the unreactive whatever the stack emitted.

Life and disease. Transmission chains preferentially carry lineages with a higher chance of onward transmission, and that chance is under no obligation to track how sick anyone becomes, so a downstream sample of infections over-represents transmissible variants rather than virulent ones. A bottleneck at each transmission event passes only part of a within-host population, so composition ratchets across successive hosts.

Information and institutions. A retold story keeps the clauses that are easy to retell; a forwarded corpus consists of whatever was forwardable; a report climbing a hierarchy clears each level with a probability keyed to how it reads on arrival; the disputes that reach an appellate court are the ones worth appealing rather than a sample of disputes. In each case the far-end record is the intersection of what existed with what could travel. ^[4]

Engineered channels. Lossy links shed traffic by size or class, relays and caches retain what gets requested again, and every intermediate hop operating a policy is one more place where the mixture changes shape.

Clarity

The abstraction pulls apart two questions that ordinary talk runs together: what is out there, and what gets here. Once they are separated the channel becomes an object in its own right, with parameters to be measured rather than a transparency to be assumed, and a disagreement between a near measurement and a far one stops being a discrepancy to be resolved and becomes a datum about the path between them. ^[5]

It dissolves a second confusion, the slide from prevalence to merit. That some idea, product, strain or story is what one keeps encountering is evidence about the encountering; the further claim that it is also good, true or important must rest on other grounds. The prime does not say prevalence is uninformative. It says prevalence reports the joint effect of source and channel, so any statement about the source alone requires the channel's contribution to be estimated and subtracted first.

Manages Complexity

What the abstraction lets an analyst stop tracking is the interior of the path. Individual histories, the identity of each relay, the microstructure of the medium, and the particular reason any given casualty was lost all collapse into a single number per component — a coefficient per unit distance, or a retention probability per hop — from which the mixture at any point follows. A forty-hop path needs no more parameters than a four-hop one. ^[1]

It also retires a shelf of parallel vocabularies. Absorption, scattering, sedimentation, immune clearance, packet drop, editorial rejection and plain forgetting all occupy the same slot in the same rule, so a technique developed against one of them is available against the rest without being reinvented under a local name.

Abstract Reasoning

The prime licenses a compact diagnostic, and the first step is the counterfactual that defines it: would this item have been as likely to arrive here had its read property been different, everything else held fixed? An affirmative says the path was indifferent to the attribute; a negative says the observed population is a survivor set and has to be handled as one. ^[6]

The second step is to measure a gradient instead of a level. Sample the population at two or more points along the path, take the logarithm of the ratio between any two components at each point, and use the change in that quantity between stations to estimate selectivity — an estimate in which the unknown source composition cancels out of the arithmetic entirely. A flat gradient reports uniform loss however heavy the total attrition; a steep one reports strong selectivity however light. ^[7]

The third step separates loss from lateness. At any finite observation horizon a missing component may be gone or may merely be slow, the two call for opposite responses, and the transit time therefore has to be bounded before survival is concluded from an absence.

The fourth is correction rather than despair. Where per-component passage probability can be estimated, downstream counts can be reweighted by its reciprocal to recover an estimate of the input, subject to a standing caveat: classes whose passage probability approaches zero contribute no observations to reweight, and no weighting scheme recovers what left no trace.

Knowledge Transfer

What travels between substrates, on this entry's account, is the algebra and the inference procedure built on top of it — survival composing multiplicatively along a path,^[8] the exponential-in-distance divergence that follows in an absorbing medium,^[1] the geometric-in-hops divergence the same product rule gives across identical hops, the cancellation that makes selectivity estimable without ever seeing the source, and the fork between repairing the channel and repairing the input. This entry reads an engineer equalising an optical link and an editor redesigning an escalation form as applying one result in two vocabularies. ^[9]

What does not travel is everything about reversibility, observability and reflexivity in the particular medium. An absorbed photon is gone and cannot be reinjected further along, whereas a suppressed report can re-enter the chain by another route. A transmitter's launched spectrum can be measured directly, whereas the repertoire of an extinct culture cannot. Absorption coefficients in glass do not change because light has already passed through, whereas immunity, saturation and audience fatigue mean that passage itself rewrites the survival rule in biological and social media. Above all, components in informational substrates can adjust the very attribute the channel reads, which makes the rule adversarial in a way that no physical attenuation coefficient ever is.

Examples

Formal/abstract

Take a beam carrying two components with initial amounts N₁(0) and N₂(0) and attenuation coefficients α₁ < α₂ per unit distance, so that Nᵢ(x) = Nᵢ(0)·exp(−αᵢx). The ratio of the two then evolves as ln[N₁(x)/N₂(x)] = ln[N₁(0)/N₂(0)] + (α₂ − α₁)x, and three consequences follow immediately. The composition shift is exactly linear in distance in log-ratio coordinates whatever the absolute intensity is doing, so a signal can be fading quickly and shifting slowly, or the reverse. The shift depends only on the difference between the coefficients, so a path removing 99.9% of everything at equal rates produces no shift at all, while a nearly lossless path with a small differential produces a large one given enough distance. And because the source term is a constant, subtracting the log-ratio measured at one station from the log-ratio measured at another returns (α₂ − α₁) directly, with the initial amounts cancelling. The discrete version substitutes pᵢ raised to the number of hops k, so the ratio between components grows as (p₁/p₂)ᵏ and a per-hop advantage of a few percent becomes an order of magnitude after enough relays. In both versions the mixture converges as x or k grows on whichever component the path treats best. ^[1]

Mapped back: the reddening sunbeam, the hardening X-ray beam, the smoothed distant seismogram and the story that has shed every clause hard to repeat are one equation worked in four substrates. Linearity says the shift is a dose measured in path length rather than an event. The difference-of-coefficients result says severity and selectivity are independent axes and must be measured separately, which is why a channel's reputation for being lossy tells you nothing about whether its output can be read at face value. And the cancellation says an analyst can characterise a channel from two downstream positions without ever obtaining access to the source, which is the only reason far-field correction is possible at all.

Applied/industry

An operator with four management layers between the shop floor and the executive committee wants to know why serious problems keep arriving as surprises. No one is lying, and every layer forwards most of what it receives. But forwarding is not indifferent to content: an item that is unambiguous and cheap to own moves up readily, while one that is ambiguous, expensive, or likely to attract blame for the person forwarding it moves up less readily. Suppose those two classes clear each layer with probabilities of 0.9 and 0.6 respectively — a single layer's difference is barely perceptible, and any manager auditing their own inbox would find the chain broadly honest. Across four layers the ratio between the classes changes by (0.9/0.6)⁴, roughly a factor of five, so the committee's picture is about five times more weighted toward the comfortable class than the floor's, produced entirely by small local reluctances that no one could point at. Three remedies are available: shorten the chain through skip-level review and site visits, add a parallel route whose survival rule differs through anonymous reporting or an audit function that reports elsewhere, or precondition the content through structured forms that strip the penalised features before any layer can penalise them. ^[10]

Mapped back: the medium is a reporting hierarchy, the read property is how an item looks to whoever decides whether to pass it on, and the arithmetic is the hop form of the attenuation law. The committee's picture is a survivor set. Because the divergence is geometric in the number of layers, removing a layer buys more than exhorting any layer to try harder. And the correction requires forwarding rates that the top of the chain cannot observe from where it sits, which is why the parallel route and the direct sample are usually cheaper than the estimate.

Structural Tensions

T1 — The far-end census is partly a fact about the channel. Every downstream measurement answers a compound question — what was emitted, and what could reach here — while almost always being reported as an answer to the first alone. The temptation is worst where the data look best: a long, clean, high-volume record from a single position is exactly the record whose selectivity is invisible, because nothing in it varies. Additional depth there only tightens confidence intervals around a shifted quantity. The discipline the prime imposes is to spend part of the measurement budget on a second position instead.

T2 — Selectivity compounds, so long paths speak in one voice. Because passage probabilities multiply, per-hop differences too small to notice locally become dominant globally, and every actor along the way can truthfully report having passed on nearly everything they received. Diversity observed at the far end therefore understates diversity at the source by a factor that grows with relay count. Auditing any single hop finds nothing wrong, because at any single hop nothing is wrong; the defect is a property of the chain's length that no participant in it is positioned to see.

T3 — What travels well and what is worth having are separate variables. The read property is under no obligation to correlate with accuracy, importance or fitness, so a component can dominate the far end while being the least valuable thing that set out. Worse, the attributes that assist transmission — brevity, simplicity, extremity, low cost to pass along — are frequently the ones that fidelity must spend to obtain. Where the two trade against each other the mixture's transmissibility rises while its informativeness falls, and the far end grows steadily more confident and less right.

T4 — Two places to intervene, and the reachable one is corruptible. Selectivity can be attacked at the channel by flattening the survival rule, or at the content by changing what is launched so that survivors come out in the wanted proportions. Channel work is general and durable and usually expensive or outside anyone's authority. Content work is cheap and immediate, and it also teaches every participant precisely which attribute the channel rewards. Once that is common knowledge the attribute is optimised directly, whatever it stood for decays away from it, and the channel goes on selecting on a signal that has stopped meaning anything.

T5 — Absent and not-yet-arrived are the same observation. Selection acts over distance and distance takes time, so at any finite horizon the missing components are a mixture of the genuinely lost and the merely slow. Reading a partial record as a final one overstates selectivity; waiting for a completeness that never comes forfeits the ability to act. The asymmetry bites hardest when the slow components are also the informative ones — the careful investigation, the long-latency failure, the rare arrival — so an early picture is not merely incomplete but shifted in a direction one can predict in advance.

T6 — Correction demands the rule, and the hardest-hit classes leave no trace. Reweighting survivors by the reciprocal of their passage probability recovers the source only if those probabilities are known, and they are ordinarily estimated from the survivors themselves, which is the very population the estimate is meant to repair. Classes whose passage probability is effectively zero contribute nothing to any such estimate; they are missing from the record and from its correction alike. Aggressive reweighting of a handful of survivors then manufactures a confident reconstruction that fails more quietly than no correction would have.

Structural–Framed Character

Selective Propagation sits at the structural end of the structural–framed spectrum at an aggregate of 0.0, every criterion reading exactly zero. What travels is a transit-time selection rule: a heterogeneous population enters a common path, carrier or medium; each component has a property changing its probability or distance of continued propagation; components are preferentially retained, blocked, settled, absorbed or attenuated along the way rather than assigned destinations at launch; and the population observed farther along is compositionally shifted. Continued reach, not speed, discriminates.

Nothing pulls this one off zero. The criterion that ordinarily would, human-practice-bound, reads 0.0 because the paradigm cases need no agent: optical modes, seismic waves, pathogens and chemical species instantiate property-dependent survival in transit, and messages in networks and traits in transmission chains join them unchanged.

The rest read zero for the same kind of reason. Domain vocabulary is 0.0 — path, attenuation and retention are generic, and the selection rule must name its own controlling property rather than borrow a field's. Evaluative weight is 0.0: a compositional shift is not a loss or an improvement. Institutional origin is 0.0: no discipline's apparatus is needed for differential survival along a path. Import-vs-recognize is 0.0 — a seismologist and a network engineer each recognize the same mechanism in place.

With a clean zero the prime travels wherever the roles survive. The only discipline it imposes is discrimination from its neighbours: differential transport changes velocities, filtration usually invokes a localized barrier, and the broader selection relation lacks ordered distance. Confirm progressive travel and cumulative exposure before applying it.

Substrate Independence

Selective Propagation is a highly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. What recurs is a mixed input sent down one shared path, a rule mapping some read property to the chance of continuing, attrition compounding with distance or hops, and an output whose composition has shifted although no survivor was altered. Wavelengths in a fibre, seismic energy through rock, particles settling from a suspension, reactive species crossing a column, pathogen variants along transmission chains and claims retold through relays all fit. Its limit is a commitment the formula keeps: it presumes progressive travel through a common medium that accumulates exposure, so a system sorting once at the entrance, however selectively, falls outside rather than being a further instance.

• Composite substrate independence — 4 / 5
• Domain breadth — 4 / 5
• Structural abstraction — 4 / 5
• Transfer evidence — 4 / 5

Relationships to Other Abstractions

Current abstraction Selective Propagation Prime

Parents (1) — more general patterns this builds on

• Selective Propagation is a kind of Propagation Prime

  Selective propagation is a propagation subtype that adds attribute-dependent survival and downstream composition shift to ordinary source-through-medium spread.

Children (2) — more specific cases that build on this

• Nuclear export signal Domain-specific is a kind of Selective Propagation

  The proposed strict upward parent is prime:selective_propagation.

• Sediment Transport Domain-specific is a kind of Selective Propagation

  Sediment transport is the granular-fluid specialization in which size, density, and settling response control continued mobility and therefore sort the transported population coarse-to-fine along the energy gradient.

Hierarchy path (1) — routes to 1 parentless root

• Selective Propagation → Propagation

Neighborhood in Abstraction Space

Selective Propagation sits in a moderately populated region (50^th percentile for distinctiveness): it has near-neighbors but no dense thicket of synonyms.

Family — Trajectories, Thresholds & Path Dependence (39 primes)

Nearest neighbors

• Propagation — 0.76
• Dispersion — 0.71
• Response-vs-Propagation Race — 0.71
• Selection Vs Transmission Decomposition — 0.71
• Exposure Pathway — 0.70

Computed from structural-signature embeddings · 2026-09-10

Not to Be Confused With

Selective propagation is a species of Propagation, and it differs by exactly one commitment. Propagation names the systematic spreading of a signal, effect or state from a source through a medium whose structure governs speed, attenuation and the routes available; it is satisfied by a single homogeneous thing moving outward, and its characteristic questions are how fast and how far. Selective propagation additionally requires the travelling population to be heterogeneous and the medium to respond differently to its parts, so that reach itself becomes attribute-dependent. Every instance of the child is an instance of the parent. The parent becomes the child at the moment one mixture goes in and a differently proportioned mixture comes out.

Selection is the broader retention relation: out of an available population, some criterion, pressure or rule grants certain alternatives greater retention, passage or weight, yielding a survivor set or a shifted composition. Selective propagation is what selection becomes when that criterion is applied continuously along an ordered path with cumulative exposure. Selection is satisfied by a single event — one screen, one round, one editor — where this prime needs extent, so that the magnitude of the shift is a function of how far the population travelled and the same rule applied twice does twice the work. Selection Bias is narrower again, and epistemic rather than physical: it names the distortion in an inference drawn from a skewed sample. That is a condition one may suffer as a result of selective propagation rather than the mechanism itself; the mechanism runs identically when nobody is drawing any inference at all.

Gatekeeping and Structural Filtering both localise the selectivity in a way this prime does not require. Gatekeeping puts an actor or mechanism at a choke point exercising selective passage control that the audience cannot directly observe; it presumes a point rather than an extent, and normally a party with discretion. Structural filtering describes output passing through several parallel institutional filters so that what survives is their intersection — again a finite set of nameable stages, with an institutional frame supplying them. Selective propagation covers both as special cases and extends to what they exclude: smooth attenuation with no stage anywhere, no discretion, and nothing to point at as the filter. Where a gate can be named, naming it is more informative; where the loss is spread evenly across a hundred kilometres of rock, gatekeeping has nothing to attach to.

Diffusion and Contagion describe the spreading and stay silent about the mixture. Diffusion is spread over time; contagion is the spread of a state from element to element through contact. Both are entirely compatible with a homogeneous item, and neither requires that anything be lost preferentially. Selective propagation can ride on either — a contagious process with strain-dependent transmission probabilities is the epidemiological instance — but a diffusion or contagion in which every item spreads alike is not an instance of this prime however far it reaches. What settles the question is whether the arriving mixture differs from the departing one; the fact of spreading does not settle it.

Amplification increases a signal or a disturbance, and differential amplification can certainly produce a composition shift, which makes the neighbourhood a real one. The separation is in what each pattern is about. Amplification concerns magnitude and is fully specified by a gain applied to a signal that may consist of a single component; selective propagation concerns proportion and is not even stated until two components are compared. A path can amplify everything tenfold and be perfectly non-selective, or attenuate everything and be strongly selective. Where a channel does boost some components more than others both primes apply, answering different questions: how loud, and made of what.

Selection vs. Transmission Decomposition is the accounting identity splitting a change in a population's weighted mean into a selection term, from differential weighting of units, and a transmission term, from units changing internally. It is the natural instrument for establishing that a case belongs here, because selective propagation lives entirely in the selection term: survivors are not modified in transit. A shift whose decomposition loads on the transmission term is a story about components changing along the way rather than about which ones got through, and it calls for a different investigation. Percolation sits one question upstream of all of this, asking whether a connected path exists at all once local links cross a critical density, rather than which components survive a path that already exists.

Solution Archetypes

No catalogued solution archetypes reference this prime yet.

Notes

Two modelling choices shelter under the one abstraction and repay separating in any application. Passage may be genuinely stochastic per unit of exposure, in which case a component's fate is a draw and the population thins smoothly, or it may be a deterministic threshold on the read property, in which case the population is cleanly truncated and nothing past the cut survives any distance whatever. The resulting composition shifts resemble each other over a short path and diverge sharply over a long one, and they call for different estimators. The abstraction does not choose between them; an application must.

References

[1] Beer, August. "Bestimmung der Absorption des rothen Lichts in farbigen Fluessigkeiten". Annalen der Physik, 1852. Establishes exponential attenuation with a single component-specific coefficient per unit path length, so a mixture's composition shift accumulates with distance covered while the surviving components are themselves unaltered. registry ↩a ↩b ↩c ↩d

[2] Rayleigh, Lord (J. W. Strutt). "On the Light from the Sky, Its Polarization and Colour". Philosophical Magazine, 1871. Establishes strongly wavelength-dependent scattering, so a transmitted beam reddens as path length grows and a low sun is seen through a reddened, not a defective, sky. registry ↩a ↩b

[3] Vosoughi, Soroush, Deb Roy, and Sinan Aral. "The Spread of True and False News Online". Science, 2018. Shows that what a diffusion channel carries farthest does not track the accuracy of what is carried, so the surviving attribute and merit come apart. registry ↩

[4] Bartlett, Frederic C. Remembering: A Study in Experimental and Social Psychology. Cambridge University Press, 1932. Supports serial reproduction as a channel that retains the easily transmissible material and loses the rest, so the far-end record is not a sample of the source. registry ↩

[5] Shannon, Claude E. "A Mathematical Theory of Communication". Bell System Technical Journal, 1948. Supports treating source and channel as separate objects with independently specifiable characteristics rather than reading a received message as the source's own. registry ↩

[6] Hernan, Miguel A., Sonia Hernandez-Diaz, and James M. Robins. "A Structural Approach to Selection Bias". Epidemiology, 2004. Defines selection structurally, so that a population counts as a survivor set exactly when the probability of appearing in it depends on the attribute under study. registry ↩

[7] Aitchison, John. The Statistical Analysis of Compositional Data. Chapman & Hall, 1986. Supports log-ratio coordinates as the analysis of compositions in which a common multiplicative source term cancels between stations. registry ↩

[8] NIST/SEMATECH. e-Handbook of Statistical Methods, §8.1.8.2, "Series Model." Supports only the algebra of series composition — "R_S(t) = product of R_i(t)" — that survival along a chain is the product of the per-stage survivals. The geometric decline in hop count is arithmetic from that product when the stages are identical, as the prose now says, and the handbook treats engineered series systems only, not social or informational chains. registry ↩

[9] International Telecommunication Union. ITU-T Recommendation G.650.1 (07/2010), cl. 3.6.2 and 5.4.1.3.2. Supports only the optical half of the pairing: attenuation on a fibre link is defined as a logarithmic ratio of input to output power, "A(lambda) = 10log(P1(lambda)/P2(lambda)) dB". It says nothing about escalation forms or editorial work, and reading the two tasks as one result in two vocabularies is this entry's own transfer thesis, not the Recommendation's. registry ↩

[10] Rosen, Sidney, and Abraham Tesser. "On Reluctance to Communicate Undesirable Information: The MUM Effect". Sociometry, 1970. Supports content-dependent forwarding probability in which unwelcome items clear each relay less readily than welcome ones. registry ↩