← Research
Financial Crime Detection
Spectral forensics — the Fiedler vector that finds fraud also finds disease mutations and places chip gates
JIM’S OVERSIMPLIFICATION
Normal money flows through legitimate networks. Fraud creates connections where there shouldn’t be any — coupling that doesn’t belong. There is one equation that finds the bottleneck in any network — money, protein, chip — and the fraudulent structure always sits at the bottleneck. Same equation. Same code. Different inputs.
K IN THIS DOMAIN
K here is transaction coupling. Normal transactions couple through legitimate networks. Fraud creates anomalous edges — coupling where there shouldn’t be any. The spectral gap detects the anomaly.
Normal money flows in one direction. You pay your employees. Your customers pay you. Clean, simple, mostly one-way.
Fraud money goes in circles. Enron sent money to shell companies that sent it back as “revenue.” Madoff took new investor money and sent it to old investors as “returns.” FTX shuttled customer deposits to Alameda and back.
So we built a tool that reads the shape of money flow. Not the amounts. Not the names. The shape. We tested it on five famous frauds: Enron, Madoff, Wirecard, FTX, Danske Bank. 5 for 5.
Then we discovered something unexpected: the same equation that catches fraud also finds disease-causing mutations in proteins (0.82 AUC) and places 40 million transistors on a chip in 4.5 seconds. A bottleneck in money flow and a bottleneck in protein structure are the same mathematical object: a sign change in the second-smallest eigenvector of the Laplacian.
The equation is from 1973 (Fiedler). We did not discover it. We ran it on three things nobody thought to run it on at the same time.
THE EQUATION
L = D - A
L = graph Laplacian
D = degree matrix (diagonal: how connected each node is)
A = adjacency matrix (who connects to whom)
Second-smallest eigenvector of L = Fiedler vector
It splits ANY graph into two communities. Finds bottlenecks.
The nodes don’t matter. The coupling does.
THE METHOD
1. Build a directed graph from transaction data
Nodes = entities. Edges = money flows.
2. Compute the graph Laplacian and its eigenvectors
Small eigenvalues reveal clusters and circular flows
3. Score risk factors:
• Circular flow — money that returns to its origin
• Bidirectional matching — A pays B, B pays A
• Concentration risk — too much through one node
• Layering depth — chains of shells
4. Aggregate: 0–25 clean
26–50 borderline
51+ fraud signal
ENRON SPE NETWORK
Risk: 70
Enron → SPE1 (debt transfer) → SPE1 ↔ SPE2 (cross-guarantees) → SPE1 → Enron (fictional revenue)
Spectral signature: circular flow + bidirectional matching
MADOFF PONZI SCHEME
Risk: 75
Investors → Fund → Investors (“returns” from deposits, not earnings)
Spectral signature: extreme concentration + circular flow
Fiedler vector puts Madoff Fund at the center — the geometry of a Ponzi.
WIRECARD ROUND-TRIP
Risk: 60
Wirecard → Processor A → Wirecard (same amounts out and back)
Spectral signature: bidirectional matching + circular flow
FTX / ALAMEDA
Risk: 89 — highest in the test set
Customers → FTX → Alameda → FTX (paper returns) + Alameda → Real Estate, Ventures, Donations
Every red flag at once. The FTX↔Alameda pair dominates the first eigenvector. Outward flows to non-returning sinks create spectral asymmetry that screams extraction.
DANSKE BANK LAYERING
Risk: 59
Russia → Shell A → Shell B → Shell C → Danske Estonia → EU banking system
$230 billion over 8 years. Lowest fraud score — layering is hardest to catch. Spectral method catches it through chain topology, not individual transactions.
CLEAN CONTROLS
Payroll: Risk 7 — clean. One-directional fan-out.
Supply Chain: Risk 33 — borderline. Legitimate circular flow (returns, rebates). This is the one false positive.
VC Fund: Risk 7 — clean. Long time horizons between flows.
THREE DOMAINS, SAME CODE
1. Financial transactions
Nodes = accounts. Edges = money flows.
Fiedler vector finds: circular flows, wash trading, layered structures.
Result:
5/5 fraud patterns detected
2. Protein contact maps (
Mutation Scanner)
Nodes = amino acids. Edges = physical contacts (<8Å).
Fiedler vector finds: structural bottlenecks, damage propagation paths.
Result:
0.82 AUC as single structural feature for pathogenicity.
3. Circuit netlists (
Chip Fast)
Nodes = gates. Edges = wires.
Fiedler vector finds: critical paths, placement bottlenecks.
Result:
40M gates placed in 4.5 seconds
A bottleneck in a money flow and a bottleneck in a protein backbone and a bottleneck in a chip layout are all the same mathematical object: a sign change in the Fiedler vector. One number. Same code path in all three domains.
LIMITATIONS
What the model catches: Structural fraud — money flow topology that cannot exist in honest commerce. Circular flows where value doesn’t grow. Concentration where diversification should exist.
What the model misses: Insider trading (no unusual flow). Embezzlement from a single account (no graph). Price manipulation (requires market data). Structure is all it sees.
The supply chain false positive (risk 33, borderline) is an honest limitation. Legitimate businesses with circular money flow can trigger the detector. A human reviewer clears it quickly.
Cross-domain note: We are not claiming proteins ARE financial networks. The analogy is structural (both are graphs with coupling), not mechanistic (the physics differ). Fiedler 1973, Cheeger inequality — well-known spectral graph theory.
COMPUTATION DETAILS
Machine: Mac Mini M4 · $499 · 35W
Analysis: Laplacian eigenvector decomposition on directed transaction graphs
Fraud patterns tested: 5. Clean controls: 3.
Overall accuracy: 7/8 (87.5%)
Fraud detection: 5/5 (100%)
Clean pass: 2/3 (1 borderline false positive)
Package: pip install begump. Graph Laplacian math, not neural network.
This is computational research, not legal or financial advice. Transaction patterns flagged may have legitimate explanations.