How Science Works | How Threshold Connections Turn Separate Pieces into a Network

Imagine a flat square grid of junctions. Each junction can connect to its nearest neighbours above, below, left and right. A possible connection is called a bond. In our first model each bond is independently open with probability p, and closed with probability 1 − p. Open means passable according to the model’s rule; it does not mean an electrical switch is physically open.

All junctions remain present. This is bond percolation. A different model removes junctions themselves, producing site percolation. Removing one site can remove several routes simultaneously. The two models need not have the same threshold, even on the same underlying grid.

A cluster is a set of sites connected through open bonds. Two sites can belong to the same cluster without sharing a direct bond: a route may wind through many intermediate sites. An isolated junction is also a cluster, containing one site.

For a finite sample, define the target before looking at the result. We will ask whether any open path joins the left boundary to the right boundary. Reaching the top instead does not count. Changing the sample’s aspect ratio or the boundary rule changes this event.

In a porous material, sites might represent pore bodies and bonds their connecting throats. That representation is a modelling decision requiring evidence. Pore volume alone does not specify throat connectivity, and a geometrically connected throat need not admit a fluid under every pressure or wetting condition. The simple graph isolates one part of the physical problem.

Contents · Next chapter

For a reproducible diagram, draw nine dots in three rows. Label the top row A, B, C; the middle row D, E, F; and the bottom row G, H, I. Draw all twelve possible horizontal and vertical nearest-neighbour bonds faintly. The left boundary contains A, D and G; the right contains C, F and I.

In sample X, darken all six vertical bonds and no horizontal ones. Half the possible bonds are open. Each column forms a separate cluster, but no cluster reaches from left to right. There is no crossing.

In sample Y, darken all six horizontal bonds and no vertical ones. Again half the bonds are open. Now each row is a complete left-to-right route. There are three separate spanning clusters. A crossing exists even though not every site belongs to one common cluster.

The comparison holds the number of open bonds fixed while changing their arrangement. It demonstrates why an occupancy fraction is insufficient to predict the result for an individual sample. It also separates “there is at least one crossing” from “the whole sample is connected”.

Neither drawing is a typical random sample chosen to represent p = 0.5. They are deliberately selected configurations. Under independent opening at p = 0.5, each particular twelve-bond arrangement has probability 1/4096. Showing one convenient drawing cannot establish the frequency of crossing across the full ensemble. For that we must count or repeatedly sample configurations.

Contents · Previous chapter · Next chapter

Twelve bonds, each with two possible states, give 2¹² = 4096 configurations. At p = 0.5 all are equally probable. We enumerated every configuration for this original teaching example and tested whether a path connects any left-boundary site to any right-boundary site. Exactly 2752 configurations cross, giving a crossing probability of 2752/4096 = 0.671875.

A path-finding check is straightforward. Mark A, D and G as reachable. Follow every open bond out of a marked site and mark any newly reached site. Repeat until nothing changes. The sample crosses if C, F or I has been marked. This avoids accidentally counting a diagonal touch, a closed bond, or a route that stops one junction short.

Away from p = 0.5, configurations are not equally weighted. A specific arrangement with k open bonds has probability pᵏ(1 − p)¹²⁻ᵏ. Let Nₖ count crossing configurations containing exactly k open bonds. Then the exact crossing probability is the sum of Nₖpᵏ(1 − p)¹²⁻ᵏ over k from zero to twelve.

For checking an implementation, the thirteen Nₖ values in that order are 0, 0, 3, 34, 162, 422, 671, 690, 473, 218, 66, 12 and 1. Their sum is 2752. The three two-bond crossings are precisely the three complete horizontal rows.

Weighting this enumeration gives crossing probabilities about 0.13591, 0.48692, 0.67188, 0.82301 and 0.97890 at p = 0.2, 0.4, 0.5, 0.6 and 0.8. These are exact finite-model calculations rounded for display, not measurements on a physical material.

Contents · Previous chapter · Next chapter

Our small grid crosses with probability about 67% at p = 0.5. That does not contradict the famous threshold of one half for independent nearest-neighbour bond percolation on the infinite square lattice. The two statements concern different events.

For the infinite model, define θ(p) as the probability that a specified site belongs to an infinite open cluster. The critical value marks the onset of positive θ. Kesten’s original 1980 paper proved that this critical value is exactly 1/2. For this model θ is zero at and below one half, and positive above it. A finite crossing at the critical value can still occur: a long cluster need not be infinite.

No finite twelve-bond system has an infinite cluster. Its crossing probability is a smooth polynomial in p, rather than a sudden mathematical jump. As increasingly large samples are considered, crossing behaviour can become sharply sensitive to occupancy near the infinite-system threshold.

The value one half belongs to the specified lattice, bond rule and independence assumptions. It is not a universal statement that every conducting mixture starts working when half its volume is conductive. Long rods, three-dimensional pore networks, directional links and correlated placement change the connectivity problem.

There is another distinction worth retaining. Even above the threshold, a chosen site is not guaranteed to belong to the infinite cluster. The statement that some infinite cluster exists and the probability that a particular site joins it are related but different questions. Finite experiments require their own boundary-defined observables.

Contents · Previous chapter · Next chapter

An independent random trial redraws all twelve bonds using the same opening probability. Counting the fraction of trials that cross estimates the finite crossing probability. This answers a different question from opening bonds one at a time in a single sample until its first crossing appears.

At p = 0.5, the exact crossing probability in our model is 0.671875. With 100 independent trials, the standard deviation of the measured crossing fraction is the square root of R(1 − R)/100, where R is that probability. It is about 0.04695, or 4.7 percentage points. A result several percentage points from 67.2% is therefore unsurprising.

Increasing the number of independent samples reduces sampling uncertainty; enlarging each grid changes the underlying finite-size problem. These are different improvements. One very large grid is not automatically a substitute for many independently prepared grids.

If the same underlying random numbers are used to compare several p values, the samples across p are deliberately paired. That can make a monotonic comparison clearer, but those observations should not be analysed as independent replications. Similarly, repeated photographs of one film do not create new independently manufactured films.

A useful report records grid size, bond or site rule, direction of crossing, boundary conditions, trial count and randomisation method. It distinguishes a predicted curve from sampled points and supplies uncertainty for the latter. A beautifully smooth curve drawn through a few small samples does not make the estimated threshold exact.

Contents · Previous chapter · Next chapter

Now leave the nine-dot counting problem and consider a separate electrical toy model. A route contains three ordinary resistors in series, each 100 ohms. The route exists, and its total resistance is 300 ohms. Two such routes in parallel, sharing only their end terminals, give 150 ohms.

Replace the middle resistor of the single route with a 10,000-ohm junction. The graph is still connected, but its resistance becomes 10,200 ohms. At a fixed small applied voltage in this ideal linear model, it carries 34 times less current than the original 300-ohm route. The yes-or-no connectivity test has not changed at all.

This example assumes finite positive ohmic resistances, ideal terminals and no other paths. Under those assumptions a connecting path gives nonzero terminal conductance. It does not specify whether that conductance is useful, measurable above noise, or stable under operation.

Physical networks add junction barriers, contact resistance, changing geometry and sometimes nonlinear conduction. A microscopy image can suggest that particles touch while missing an electrically important surface layer. Conversely, tunnelling can allow conduction across a small gap that a strict touching rule would label disconnected.

The 2024 original study of junction resistance in nano-networks explicitly treats particle and junction contributions in interpreting transport. Its relevance here is limited but important: knowing the arrangement of pieces is not enough to determine their electrical response. A geometrical threshold and an instrument-defined conductivity threshold need not coincide perfectly.

Contents · Previous chapter · Next chapter

A serious test combines structural information with transport measurements. Images can estimate particle positions, orientations and candidate contacts. Electrical measurements can test whether a path carries current between specified electrodes. Neither observable alone fully identifies the mechanism.

A 2016 silver-nanowire study investigated how rod orientation affects electrical anisotropy in highly transparent networks, using experiment alongside network modelling. Alignment changed the relationship between direction and conduction. The example warns against replacing a directional network by a single scalar occupancy number without checking what information is lost.

One could compare independently prepared films across a range of wire densities while recording orientation distributions, dimensions and contact treatment. If density and junction quality change together, a conductivity increase cannot automatically be attributed to new connectivity alone. A control that changes junction resistance while approximately preserving geometry helps separate the possibilities.

In a porous medium, a corresponding programme would compare three-dimensional pore connections with pressure-dependent flow, rather than infer permeability from porosity alone. Dead-end pores contribute volume but do not necessarily support through-flow. Wetting and trapped phases can further change which geometrical connections are available to the moving fluid.

These are evidence-design principles, not instructions for preparing nanomaterials, applying hazardous voltages or conducting pressure experiments. The safe reader activity is the graph calculation. Real material experiments require suitable equipment, trained supervision and a physical model appropriate to the sample.

Contents · Previous chapter · Next chapter

The basic model flips an independent coin for every bond. Real cracks, fibres or pores often have spatial correlations. A manufacturing process may produce neighbouring contacts together, while damage may remove an extended band of connections at once.

Return to the two six-bond drawings. A process that strongly favours complete vertical columns and one that favours complete horizontal rows can have the same mean occupancy but very different left-to-right crossing probabilities. The marginal probability that one bond is open does not describe the joint pattern of many bonds.

A useful diagnostic compares observed clustering or orientation with the patterns produced by the independent model. If neighbouring links tend to appear or fail together more often than expected, simulations should incorporate that dependence before their thresholds are interpreted physically.

Direction also matters. A route passable left-to-right need not be passable in reverse if links have directional rules. A network that changes over time may admit a sequence of contacts even when no instantaneous picture contains a complete route. Those are distinct model extensions, not contradictions of the static undirected calculation.

Adding open bonds to one fixed undirected graph cannot destroy an existing path. This monotonicity gives a useful implementation check. It does not imply that every physical performance measure improves when more material is added: transparency, flexibility, contact chemistry or other properties may change. The connectivity theorem only governs the defined path event.

Contents · Previous chapter · Next chapter

First ask: what is the control variable? The probability p of an independent model is not automatically a measured volume fraction, wire number density or fluid saturation. Connecting them requires a model or calibration. Reporting “the threshold is 0.5” without naming the quantity leaves the claim incomplete.

Second ask: what is the outcome? A left-to-right crossing, membership in the largest cluster, terminal conductance and fluid permeability are different observables. They may respond near a common structural change while carrying different information. Our sample Y has three spanning clusters, illustrating why crossing alone does not determine largest-cluster membership.

Third ask: what size limit is being used? A finite sample can connect below an infinite-model threshold or fail to connect above it. Repeated measurements across several sizes help separate finite-size effects from a change in the underlying model. The geometry and boundary conditions must remain explicitly comparable.

Suppose a report finds crossings in 49 out of 100 small samples at p = 0.4. That is compatible with our exact small-grid prediction of about 48.69%, but the agreement does not prove the infinite threshold or validate a material’s contact physics. Many larger-scale or physical questions remain untested.

The strongest claim is the one matched to the actual evidence: a stated finite-model prediction passed a stated finite-model test. Extending it to a real material requires additional observations, rather than a more confident description of the same graph.

Contents · Previous chapter · Next chapter

Take the labelled nine-dot diagram and answer four questions before calculating anything. Does a vertical-only six-bond sample cross left-to-right? No. Does a horizontal-only six-bond sample cross? Yes. Must a crossing sample contain one cluster holding all nine sites? No. Does one randomly chosen sample determine the ensemble crossing probability? No.

Next reconstruct the two-bond part of the enumeration. Only a complete horizontal row reaches across using two bonds, so there are three successful configurations. This simple check catches common mistakes in boundary labels or diagonal adjacency before a larger calculation is trusted.

Finally compare the separate resistor examples. Both the 300-ohm chain and the 10,200-ohm chain contain a path. Their different currents show why connectivity and transport magnitude must be evaluated separately. Adding a second independent 300-ohm route reduces the ideal resistance to 150 ohms without changing the answer to the original yes-or-no question.

For deeper reading, start with Kesten’s threshold proof, then compare the physical questions addressed by nanowire orientation experiments and junction-resistance research. They connect mathematical structure with measurement while retaining different assumptions.

The central lesson is precise: local availability can create long-range connectivity collectively, but a threshold belongs to a defined model and observable. Continue through the Statistical Mechanics guide or return to the How Science Works master guide to place that lesson alongside other collective phenomena.

Contents · Previous chapter · Continue to Statistical Mechanics