What coupled topologies revealed in magnetospheric data — a systems-thinking worked example.
A colleague asked me to share my experience using systems thinking to solve problems. Here is one from magnetospheric physics—not because the domain matters, but because the structural pattern transfers.
I was analyzing satellite data that carried three independent structures: spatial, spectral, and temporal. Standard practice handles each one separately, projecting away the others. I kept finding features in the raw data that disappeared under single-structure analysis. So I built a minimal relational framework to couple the structures: two field types, three operators whose composition is constrained by their types and mathematical content, and a directed coupling cycle between topologies.
The framework resolved cross-topology features—field-aligned currentsElectrical currents flowing parallel to Earth's magnetic field lines, connecting the magnetosphere to the ionosphere — analogous to directed data flows along fixed network paths., injection boundariesSharp transitions where energetic plasma enters the inner magnetosphere from the tail region — like a sudden load spike propagating from one service tier to another.—that single-topology analysis could not cleanly isolate. No optimization target, loss function, or model selection was involved. This tutorial walks through the construction, the constraints that determine it, and what it revealed.
You know what a projection does to a vector space. Project ℝ3 to ℝ2 by dropping a coordinate and you lose everything the dropped coordinate carried. If coordinates are independent, no harm. If they correlate, the projection distorts the surviving structure.
This is not unique to linear algebra. Any time data has multiple natural structures and you analyze one at a time, you are projecting. Spatial analysis of a multi-energy dataset throws away spectral structure. Spectral analysis at a fixed location throws away spatial context. Temporal analysis at a fixed point discards everything about how the signal propagates.
Standard practice treats this as acceptable: analyze each structure independently, combine the results. But some features exist in the coupling between structures. They are not hidden in any single projection—they are absent from it. A spatial analysis at any resolution cannot find a feature that is defined by the relationship between spatial and spectral structure. The feature is not faint; it is structurally inaccessible to single-topology methods.
Consider a finite set of nodes connected by some adjacency relation—a graph. Data on this graph takes one of two forms, and this distinction is not a design choice. It is the minimum structure needed to separate where something is from how things relate.
A NodeField assigns one scalar value to each node. It describes the state at a location. A NodeField has no inherent directionality—the value at node i is just xi.
An EdgeField assigns one scalar value to each directed pair of adjacent nodes. It describes the relationship between locations. An EdgeField is inherently directional: the value on edge (i, j) need not equal the value on edge (j, i).
These are the only two representation types. Every quantity in the framework is either a NodeField (one value per node) or an EdgeField (one value per directed adjacency pair). Readers familiar with algebraic topology will recognize these as 0-cochains and 1-cochainsFunctions assigning values to nodes (0-cochains) or edges (1-cochains) of a graph — the formal names for NodeField and EdgeField in algebraic topology. on the graph.
The left panel shows a NodeField: four nodes with scalar values {3, 7, 5, 1}. The right panel shows an EdgeField on the same graph: each directed edge carries a signed value. Notice that opposite directions carry opposite signs: the A→B edge reads +4 while B→A reads −4. This antisymmetry is not imposed—it is a consequence of how the EdgeField was produced. I define that operator next.
How do you move from a NodeField to an EdgeField? There is exactly one natural transition, and its uniqueness follows from a single requirement: translational invarianceThe property that shifting all values by the same constant doesn't change the output — like how relative differences between metrics matter, not their absolute baseline..
For a NodeField x and each directed edge (i, j):
A(x)ij = xj − xiA extracts the pairwise difference—the relational gradient between adjacent nodes.
Why unique. Suppose we want a function F : NodeField → EdgeField that respects one structural constraint: if you shift every node value by the same constant c, the edge field should not change. Formally: F(x + c) = F(x) for all constant shifts. This is translational invariance—the absolute level of node values carries no relational information; only differences do.
For a single edge (i, j), the most general linear function of two node values is Fij = α xi + β xj. Translational invariance requires α(xi + c) + β(xj + c) = α xi + β xj for all c, which forces α + β = 0. Up to a constant scale, the only solution is xj − xi.
This is the graph coboundary operatorThe operator that maps node values to edge differences by subtracting source from target — computing the discrete analog of a gradient across a graph. d0 from discrete exterior calculusA mathematical framework extending calculus concepts (gradients, curls, divergences) to discrete structures like graphs and meshes.—the same map from 0-cochains to 1-cochains that appears in simplicial cohomology. I arrived at it from a different starting point: the pre-A admissibility constraint below. The derivation itself recovers a classical construction; what I have not seen elsewhere is the observation about which preprocessing operations it rules out.
Because A computes pairwise differences, any transformation applied to the NodeField before A must preserve pairwise differences. Which transformations qualify?
This is not a rule I imposed. It is a consequence of what A does. If you smooth the NodeField before applying A, the edge field reflects the smoothing kernel, not the data. The constraint is a theorem about the operator, not a convention about the workflow. This is the part I think is new: not A itself—which is d0—but the explicit characterization of what standard preprocessing steps do to the relational structure A extracts.
One consequence of identifying A with d0 deserves attention before we move on. The image of the coboundary map is the space of exact 1-cochains—edge fields that are pure gradients, with zero circulation around every cycle in the graph. Everything downstream of A starts in this gradient sectorThe subspace of edge fields that can be written as differences of node values — fields with no rotational component, like a DAG with no cycles.. The question for the next two operators is: which one stays there, and which one leaves?
A transforms NodeField to EdgeField. Two more operators complete the chain, and both operate within EdgeField. Their composition requires some care—because the two operators do fundamentally different things to the gradient structure that A established.
A is the only operator that accepts NodeField input. This is a genuine type constraint: there is no alternative starting point. The chain must begin with A.
B accumulates local relational context. For each directed edge (i, j), B adds to the edge value a contribution from the edges sharing its endpoints along the same axis, so that each edge comes to encode not just its own gradient but the gradient pattern of its neighborhood. Contributions are combined with the sign convention inherited from A, so that the accumulation respects edge direction: the value built up on (i, j) is the negation of the value built up on (j, i).
B preserves antisymmetry. If the input edge field is antisymmetric—e(i,j) = −e(j,i), as everything downstream of A is—then B’s output is antisymmetric as well: B(e)(i,j) = −B(e)(j,i). B does not create circulation or couple distinct axes; it amplifies and locally smooths the relational structure already present, leaving the field within the image of the coboundary map. In discrete-calculus terms, B keeps the field curl-freeHaving zero net circulation around every closed loop — a pure gradient field with no rotational structure, like a dependency graph with no cycles.: it is still, up to the accumulation, a gradient.
R introduces circulationNet flow around a closed loop in the graph — nonzero circulation means rotational structure that pure node-to-node gradients cannot produce.—the component that local accumulation cannot produce. B builds up gradient structure but leaves the field curl-free: it remains, in discrete-calculus terms, the coboundary of some node field. A pure gradient has no circulation. R adds exactly the part B cannot: a rotational component that takes the field out of the image of the coboundary map. The parameter ρ ∈ ℝ controls the strength of this added circulation and is bounded.
What R breaks. Both B and R preserve edge antisymmetry—R(e)(i,j) = −R(e)(j,i), same as B. So antisymmetry is not the symmetry that distinguishes them. The symmetry R breaks is the curl-free property: B’s output is a gradient (zero circulation around every cycle); R’s output is not. This is the precise sense in which R “breaks symmetry and creates structure”—it produces the rotational, non-gradient component of the edge field. In two or more dimensions, this is where perpendicular-axis coupling lives—R adds cross-terms that move the field out of im(d0), with ρ controlling the mixing strength. In one dimension, R introduces forward-backward directional bias that sustains dynamics under iteration, though the field remains exact on the single cycle (the loop sum is zero for any ρ). Accumulation alone can never produce the non-gradient component, because accumulation of gradients stays a gradient.
After A, we are in EdgeField. Both B and R operate there—both accept EdgeField and return EdgeField. So the ordering B→R is not forced by types. It is forced by what the operators do.
B keeps the field in the gradient sector—amplifying and extending the coboundary structure that A established. R then introduces the non-gradient component: circulation. Accumulating before coupling gives R structured input—edges that already encode local relational patterns. Coupling raw gradients (R before B) is mathematically well-defined but produces noise rather than structure: you are introducing circulation into differences that have no local context.
This is an empirical finding, not a type-theoretic proof. I tested both orderings. B→R produces sustained, physically meaningful dynamics. R→B does not.
The kernel is designated ABRCE after its five components. This tutorial focuses on A, B, and R—the core operator chain. The remaining two:
Everything above operates on a single graph—a single topology. The extension that made the difference for my work: the same set of nodes can carry multiple independent topologies, and R couples them.
In magnetospheric physics, satellite data has three natural structures:
| Topology | Nodes are... | Adjacency means... | Symbol |
|---|---|---|---|
| Spatial | Measurement locations | Nearby in physical space | S |
| Temporal | Time steps | Adjacent in time | τ |
| Component | Energy channels | Adjacent in the energy spectrum | C |
Each topology has its own NodeField and EdgeField. A and B operate within each topology independently: spatial A extracts spatial gradients; spectral A extracts spectral gradients. Nothing couples them yet.
The coupling comes from R. In the multi-topology extension, R passes edge-field structure from one topology to the next in a directed cycleA closed loop where edges follow a fixed order — here S→τ→C→S, like a unidirectional message-passing ring between three services.:
Read the arrows: spatial edge structure feeds into temporal circulation. Temporal edge structure feeds into component circulation. Component edge structure feeds back into spatial circulation. The cycle is directed—the order S → τ → C → S is a structural commitment, not a derivation. Reversing it gives different results because R is antisymmetric. The direction reflects the physical reality that spatial structure constrains temporal evolution, which constrains spectral structure, which feeds back into spatial. I chose this cycle based on domain knowledge; the framework requires that a direction be declared.
Each topology runs its own A → B pipeline independently. R is the only inter-topology operator. It takes the accumulated edge field from one topology and contributes circulation to the next.
Apply the single-topology pipeline (A → B) to a four-node graph. The NodeField values are {3, 7, 5, 1} on a diamond graph (A–B, A–C, B–C, B–D, C–D).
Earth's magnetosphere is the region of space dominated by Earth's magnetic field, shaped by the solar wind into a compressed dayside and an elongated nightside tail. Satellites orbiting through this region measure particle flux across a range of energies at each point along their trajectory. The data is a function of three variables: where (spatial position), what (particle energy), and when (time).
Spatial analysis alone (krigingA spatial interpolation technique from geostatistics that estimates values at unmeasured locations using statistical models of spatial correlation., interpolation, variogramsPlots showing how measurement similarity decreases with distance — a spatial statistics tool for characterizing correlation structure in geographically distributed data.): treats each energy channel independently. At any given energy, the spatial map shows gradients and trends. These are real but incomplete. A sharp spatial gradient at 10 keV might be part of an injection boundary or it might be a spatial artifact. Without spectral context, you cannot distinguish them.
Spectral analysis alone (energy spectra at fixed locations): shows the energy distribution at each point. A hardening spectrum (shift toward higher energies) at one location might indicate freshly injected plasma or an artifact of the satellite's orbit. Without spatial context, you cannot tell.
Temporal analysis alone: shows how values change at a fixed point. A sudden increase might be local or global, injection or compression. Without spatial and spectral context, you cannot resolve the ambiguity.
Apply the multi-topology kernel. Each topology runs its own A → B pipeline. R couples them through the directed cycle S → τ → C → S.
Two classes of structure become resolvable that single-topology methods struggle with:
Electrical currents flowing along Earth's magnetic field lines. Field-aligned currents are detectable by single-spacecraft methods—the curl-B / Ampère technique using magnetometer data is standard. What single-topology analysis struggles with is resolving which particle populations carry the current at which energies. The multi-topology coupling preserves the spectral–spatial correlation that single-instrument analysis projects away, resolving the current's spectral specificity alongside its spatial geometry.
Boundaries where plasma from the magnetotail is injected into the inner magnetosphere. Energy-time dispersions are the standard single-instrument signature. But the spatial extent and spectral profile of the boundary—features that require coupling between spatial and spectral topologies—are harder to resolve without the coupling. An injection boundary is a simultaneous change in spatial gradient AND spectral gradient: the spectrum hardens at the same spatial boundary where flux increases. This feature lives in the relationship between spatial and spectral structure. The R operator coupling C → S is what makes it cleanly resolvable.
The comparison is not just "sharper vs. blurrier." The injection boundary does not become clearer with coupling—it becomes cleanly resolvable for the first time as a cross-topology feature: defined by the simultaneous spatial and spectral gradient. Single-topology methods can detect pieces of it—a spatial gradient here, a spectral shift there—but the cross-topology coherence that defines it as a single physical structure requires the coupling to resolve.
What struck me about this approach is that I never specified what to look for. No optimization target, no loss function, no training data, no model selection, no specification of target features. The operators are determined by the type structure and the domain constraints. What emerges is what the topology admits.
This is different from goal-directed systems analysis, where you define a target state and compute paths to it—an approach that has its own strengths. The kernel is non-goal-directed: it reveals what the relational structure admits, without specifying what to find. Both approaches are valid; they answer different questions. Goal-directed analysis asks "how do I reach state X?" This framework asks "what structures does this topology support?"
The kernel is non-optimizing (no loss function), non-semantic (operators do not interpret), and non-goal-directed (no target). These are observations about what the operators do, not axioms I chose for aesthetic reasons. A is the graph coboundary—it computes differences. B accumulates local context. R introduces circulation. None of them specify what to find. The features that emerge—field-aligned currents, injection boundaries, coherence patterns—are consequences of the topology and the data.
If no cross-topology coherence exists in the data, the kernel produces none. It cannot hallucinate structure that the topology does not support.
Systems thinking, at its core, claims that structure determines behavior—that how parts relate matters more than what parts are. This framework makes that claim precise: declare the relational structure (the topologies and their adjacencies), apply the minimal type-constrained operators (the coboundary, accumulation, circulation coupling), and observe what emerges.
The recipe is simple. The hard part was trusting that less specification would produce more structure.
References & further reading.
Structure from Topology — B. Stephenson, 2026.