In the science of complex systems, a long-known problem is that pairwise connections do not tell the whole story. When several processes influence each other simultaneously, part of the information is transmitted not directly but through joint actions, or conversely, it is duplicated along different paths. Alongside this work in the archive are studies that seek higher orders in data: some dissect the echo of an invisible stone, others extract laws from noise, and still others teleport energy along a chain. All of them grapple with how to pull out not just pairwise but also collective interactions from observations. This article joins that line but shifts the focus: instead of building a complete picture of higher orders, it offers a simple and computationally light way to separate the pure two-particle contribution from everything else.
The authors start from the fact that conditional mutual information between two variables can either decrease or increase when a third is added. If it drops, the third variable duplicated part of the signal (redundancy); if it rises, the signal appears only when considered jointly (synergy). On this basis, they build a search: among all other processes, they look for a subset that maximally weakens the entropy transfer from source to target, and one that maximally strengthens it. The difference between the pairwise value and the minimum gives the redundant part, the difference between the maximum and the pairwise value gives the synergistic part, and the minimum itself is considered the unique contribution. Since an exhaustive search over all subsets is infeasible, a greedy algorithm is used: variables are added one by one until adding more no longer significantly changes the result compared to random surrogates.
On a toy model with four processes, where two sources converge at one node that then influences the target, the pairwise entropy transfer from one source to the target was 0.32, the fully conditioned transfer was zero, and the maximum was 1.25. The decomposition gave: unique contribution zero, redundant 0.32, synergistic 0.93. This exactly reflects the model's structure: there is no direct path, but information passes through the common node and through the joint action of the two sources. On real climate data, where the influence of the El Niño index (ocean surface temperature) on the Southern Oscillation (pressure difference between Tahiti and Darwin) was analyzed, the pairwise transfer was 0.078, the minimum 0.074, and the maximum 0.094. Hence the unique contribution is 0.074, redundant 0.004, and synergistic 0.016. Collective effects account for 20% of the maximum information flow.
The work aligns with neighboring papers in the archive but differs in approach: instead of a full description of higher orders, it offers a simple and fast decomposition that can be applied to large networks without huge computational costs. While neighboring articles often build complex theoretical constructs or seek new phenomena, here a practical tool is provided that works directly on data. This is the first test of the idea that searching for extreme subsets of variables can isolate a pure two-particle signal amid the noise of collective interactions. The result on climate data shows that higher orders are not exotic but constitute a noticeable fraction of the information flow.
The method opens the way to a systematic analysis of which variables, and in which combinations, create redundancy or synergy. This could lead to new ways of visualizing and understanding the structure of complex networks, where not only nodes and links but also groups of nodes matter. It might be possible to connect the proposed decomposition with more general frameworks of partial information decomposition and with the theory of higher orders that has been actively developing in recent years. It would also be interesting to test the method on other types of data, such as in neuroscience or economics, where collective effects are presumably strong.
The method applies to any time series where directed interactions need to be understood, from climatology to neurophysiology, from financial markets to social networks.
The next step could be to compare the proposed decomposition with other approaches to higher orders on extensive simulations and real data, as well as to develop statistical tests for assessing the significance of each component.
In the field of complex systems analysis, an open question remains: how to distinguish true causal links from spurious ones generated by common factors or collective effects. This work directly addresses that problem by offering a way to quantitatively estimate the contribution of collective interactions to the observed information flow.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors propose a method to decompose the information transferred between two processes into unique, redundant, and synergistic components by searching for subsets of other variables that maximally weaken or strengthen the flow. Applying a greedy algorithm with surrogate data for stopping, they demonstrate the method on a toy model and El Niño climate data, where collective effects account for about 20% of the information flow.
Where its strength is
The strength lies in its simplicity and computational efficiency: instead of a full higher-order description, the method provides a fast decomposition applicable to large networks.
Where it could go next
Worth trying
Why this may matter
Such a tool could help isolate direct causal links in complex systems, where collective effects often mask true interactions, which is important for understanding climate teleconnections and other network processes.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.57, against a usual 0.64. It looks like few have drilled nearby.
El Niño and the Southern Oscillation are so tightly linked that they are combined into the acronym ENSO, but the proposed method shows that even in such a pair, about a fifth of the information is transmitted not directly but through other climate indices, such as the temperature of the Northern Tropical Atlantic or the Southern Tropical Atlantic.