A working paper suggests that changing a system’s known actuation gains over time can make hidden links in a feedback network distinguishable from a confounding influence, even when the system’s commanded inputs and internal responses are not observed. The result is conditional: the gain changes must meet assumptions about exogeneity, signal variation and orthogonality.
The information is in the change
The study examines an “output-only” setting. Researchers see outputs before and after a clearing window, along with the actuation gains, but not the coupling response or the inputs during that window. The framework assumes that gains are exogenous, that the signals contain persistent conditional excitation and orthogonality, that the signals are stationary and ergodic with finite fourth moments, and that a partial reversal is known.
In the linearized version of the model, a changing gain path can separate a coupling from a gain-invariant confound. Constant gains cannot do that: they leave a continuum of coupling-and-confound pairs that fit the same observations. The study uses the smallest eigenvalue of a residualized information matrix as a gauge of how much usable interaction information the design contains.
The exact fixed-point analysis makes the same issue a rank question. Local identification depends on the rank of a stacked differenced Jacobian across gain regimes; when the regimes coincide, that differenced Jacobian is zero and observationally equivalent pairs remain.
A signal is not automatically a direct link
The method’s interaction coefficient is a network sensitivity that depends on the gains, not necessarily the coefficient of one direct edge. If a varying gain has an outgoing path, the analysis can identify the associated full column and row of the coupling matrix, but it guarantees only a cross-pattern of entries. The complementary block remains unresolved.
That distinction matters because a signal may travel through several links. The detector can therefore respond to a reachable indirect path as well as to a direct connection. Recovering direct topology requires a second stage, and a zero direct entry can still produce a reachable transmission signal.
What the simulations found
The main synthetic test used five channels, 750 observations and 200 Monte Carlo paths. It set the reversal share at 0.85, used three gain levels—0.5, 1.0 and 1.5—and planted four true off-diagonal couplings. The modeled spectral radius ranged from 0.30 to 0.55, with a common factor in pre-window signals and confounding assigned to pairs with zero coupling.
The results followed the identification logic, but also showed the cost of poor designs. Constant gains produced non-identification on all paths. With correlated gains, the proposed estimator had power of 1.00 and a false-alarm rate of 0.22; with independent gains, power was again 1.00 and the false-alarm rate was 0.12. When the exclusion condition was violated, RMSE rose to 0.34 from 0.28 in the baseline.
The comparisons with simpler benchmarks were less favorable. A level or Granger-style benchmark had a false-alarm rate of 0.53 when the confound mimicked a reversal. An untuned variance-ratio benchmark had an RMSE of 43.4; tuning reduced RMSE to 0.60, but 22% of paths produced no admissible estimate.
Reachability also changed the detector’s behavior. Reachable-zero flags rose from 7% when the spectral radius was about 0.15, to 39% at about 0.30 and 84% at about 0.45. The size of the unreachable-pair flags drifted to 14%, compared with a 5.3% baseline under the 5% rule. Using the full interaction regression reduced the reachable-zero rate to 27% without changing the unreachable size.
Monitoring brings a familiar trade-off
A rolling transmission monitor made the trade-off visible. In a simulation with 40 paths per arm, a looser rule—threshold z = 2.576 and two consecutive windows—flagged 90% of null paths. A stricter rule, using z = 3.72 and three consecutive windows, cut the false-alarm rate to 10%, but missed 18 of 40 treated paths. The procedure was not uniformly calibrated across edges and overlapping windows.
For a frozen network matrix, a common scalar reversal share and a real spectrum, the proposed persistence boundary is ρ(L) < 1 − θ/2. At θ = 0.88, that gives a boundary of 0.56 rather than the within-period boundary of 1. The authors describe this as a heuristic for slowly varying systems, not a general stability theorem.
The uncertainty calculations favored a design-preserving dependent-multiplier block bootstrap for the implemented two-stage pipeline. Across the reported simulation, its basic intervals covered the target in 90% of cases at a nominal 95% level, compared with 45% to 61% for the delta method and 55% for naive percentile intervals. In 100 spectral-uncertainty paths, the operational SVD slab with fallback had spectral bias of +0.098, with a standard deviation of 0.058, versus +0.131 and 0.169 for inversion alone. The SVD branch itself was not calibrated.
Financial examples are screens, not verdicts
The Korean case study applied the method to 16 single-stock leveraged exchange-traded funds linked to Samsung Electronics and SK Hynix. Intraday bars were available only after launch, so the analysis used a contaminated close-to-close regressor covering the full period. The Hynix-to-Samsung cross-reversal produced z = −2.82, or −2.72 with Newey–West errors; leaving out one day at a time gave a range from −3.05 to −2.64. The result was more negative than all 182 ordered placebo pairs, and a 10-day block-bootstrap interval excluded zero.
The U.S. panels were largely null. All six treated directions in a MicroStrategy–Bitcoin–Coinbase scale placebo had absolute z-scores no larger than 1.45. In the final 2022–2024 TSLA, NVDA and AAPL panel, the full ordered-pair interaction grid was null after multiplicity adjustment; one of 12 statistics was marginal, at t = −1.77. At the same time, gain correlations ran from 0.81 to 0.91, and the rolling gain-Gram condition number had a median of 45, with deciles from 17 to 118.
Taken together, the applications support reduced-form screening evidence rather than a clean structural estimate. The financial panels are observational, so they do not establish causal effects, and a null interaction grid does not show that every possible coupling is absent.
A method with a narrow operating envelope
The paper’s limits are closely tied to its promise. The first-order estimator can converge to a pseudo-true projection while omitted cross-interactions leave a bias that does not shrink as more observations are added. The nonlinear spectral inversion is heuristic and nonconvex, and nonnegative projection and selection can add spectral bias. Formal bootstrap validity also requires support selection and smoothness conditions that do not cover the implemented fixed-alpha rule or the SVD branch.
The document is a working paper labeled August 2026 and identified as arXiv:2608.25844v1. Its author reports an affiliation with Amazon Web Services and says the views do not represent AWS or its affiliates; no separate funding disclosure is reported.
Paper data and sources
Original title: Output-Only Identification and Spectral Monitoring of Coupled Feedback Networks with Known Time-Varying Actuation
Authors: Jihwan Woo
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text