Dynamic Causal Strength in Neural Data

• Dynamic Causal Strength (DCS) is an intervention-based, time-varying measure that quantifies how causal influence changes when a source input is perturbed in transient neural systems.
• DCS compares the observational conditional of a target with its post-intervention distribution, offering nuanced insights beyond traditional predictive measures like Transfer Entropy and Granger Causality.
• By employing structural causal models and VAR frameworks, DCS provides actionable metrics for analyzing transient events, such as sharp-wave ripples in the hippocampus.

Dynamic Causal Strength (DCS) is a time-varying, intervention-based measure of causal influence between time series, developed in the neural-signal literature for transient, peri-event, and non-stationary settings. In the cited work, DCS is a dynamic extension of the causal-strength framework of Janzing et al. to time-inhomogeneous VAR processes and their unrolled causal graphs, and in later software-oriented work it is described as a time-resolved measure grounded in Structural Causal Models (SCMs) and intervention semantics. Its central object is not predictive gain but the change in a target distribution induced by cutting or perturbing a directed causal input from a source [(Shao et al., 2022); (Janzing et al., 2012); (Nouri et al., 30 Aug 2025)].

1. Foundational idea: from causal arrows to causal strength

The point of departure is the distinction between identifying a causal arrow and quantifying how strong that arrow is. Janzing et al. argue that a causal DAG specifies which arrows exist, but not how important a given direct effect is, and they define causal influence for a set of arrows $S$ as the KL divergence between the original distribution and the distribution obtained after those arrows are cut and their open inputs are replaced by independent marginal copies: $C_S(P) := D(P\|P_S).$ For a single edge $X \to Y$, the measure can be written as an expected KL divergence between the original conditional at the target and the post-cut conditional. The conceptual interpretation is explicitly intervention-based: edges are treated as channels, and causal strength is the disturbance induced by locally corrupting those channels (Janzing et al., 2012).

DCS inherits this intervention-based logic and makes it dynamic. In the transient neural-event setting, it is introduced as a “dynamic” extension of the causal-strength idea of Janzing et al. to time-inhomogeneous VAR processes and their unrolled causal graphs. The relevant causal question becomes local in time: how much does the conditional distribution of a neural target at time $t$ change when the lagged causal input from a source is removed and replaced by an independent copy? This suggests a shift from static edge relevance in DAGs to time-indexed, event-resolved causal influence in stochastic dynamical systems (Shao et al., 2022).

A plausible implication is that DCS should be understood less as a single universal formula than as a family of intervention-based constructions that preserve the same basic semantic commitment: causal influence is quantified by comparing observational and intervened distributions, not by measuring association alone. That interpretation is consistent across the foundational causal-strength paper and the later transient-neuroscience formulations [(Janzing et al., 2012); (Nouri et al., 30 Aug 2025)].

2. Structural-causal and time-series formulations

The formal background for DCS is the SCM representation

$V_j \coloneqq f_j(PA_j, N_j), \qquad j=1,\dots,d,$

where $PA_j$ are the parents of node $V_j$, $N_j$ are exogenous noises, and the graph is acyclic. In the transient-neural setting, the target process is modeled by a time-varying VAR equation such as

$$X^1_t = \mathbf{a}^{\top}_t \boldsymbol{X}^1_{p,t} + \mathbf{b}^{\top}_t \boldsymbol{X}^2_{p,t} + \eta^1_t, \qquad \eta^1_t \sim \mathcal{N}(k^1_t,\sigma_{1,t}^2),$$

with $\boldsymbol{X}^1_{p,t}$ and $C_S(P) := D(P\|P_S).$0 denoting lagged past vectors of target and source, respectively (Shao et al., 2022).

Within this framework, DCS compares the observational conditional of the target against a post-intervention conditional in which the source past is replaced by an independent copy drawn from its marginal distribution. The cited definition is

$C_S(P) := D(P\|P_S).$1

This is an arrow-level intervention: it quantifies how much the direct mechanism into the target changes when the source-to-target link is cut, while the remainder of the model is preserved as far as possible (Shao et al., 2022).

TranCIT presents the same underlying idea in a more generic dynamic form. There, causal strength is described as a divergence between the future distribution of a target under intervention and its baseline distribution,

$C_S(P) := D(P\|P_S).$2

and, for time series,

$C_S(P) := D(P\|P_S).$3

The emphasis is that DCS is directional, local in time, and mechanistically causal rather than purely predictive. In intuitive terms, it asks: if one intervenes on the source at a particular time, how much does the target’s future change? (Nouri et al., 30 Aug 2025)

3. Relation to transfer entropy, directed information, Granger causality, and DDCM

DCS is repeatedly contrasted with predictive or dependence-based quantities. In the transient neural-event paper, Transfer Entropy (TE) is written as

$C_S(P) := D(P\|P_S).$4

or equivalently as an expected KL divergence between the full conditional and a reduced conditional that marginalizes the source past using $C_S(P) := D(P\|P_S).$5. DCS instead compares the observational conditional to an intervened conditional in which the source past is replaced by an independent copy from its marginal $C_S(P) := D(P\|P_S).$6. On this basis, TE is characterized as predictive and DCS as closer to a mechanistic intervention (Shao et al., 2022).

TranCIT sharpens the same distinction in operational terms. Granger Causality is said to ask whether past values of $C_S(P) := D(P\|P_S).$7 improve prediction of future $C_S(P) := D(P\|P_S).$8, whereas DCS asks how $C_S(P) := D(P\|P_S).$9 changes if $X \to Y$0 is perturbed. TE is described as directional and nonlinear in principle, but still fundamentally predictive or informational rather than interventional. The toolbox paper also states that TE can be inflated by common drivers, indirect pathways, finite-sample effects, and strong autocorrelation, while DCS is intended to reflect the change in the target distribution under hypothetical intervention (Nouri et al., 30 Aug 2025).

The relation to directed information is more subtle. The paper on discrete Dynamic Causal Modeling (DDCM) proves a conditional equivalence between DDCM and Directed Information (DI) for two-brain-region systems under specific assumptions: causal dynamics, Gaussian neurostate and noise, equal variances, constant external input, a discrete-time formulation, and an invertible hemodynamic mapping. Its main proposition states that if $X \to Y$1, then

$X \to Y$2

with

$X \to Y$3

That paper does not explicitly develop a formal theory of DCS, but it explicitly connects “causal-strength-type quantities” to the equivalence between model-based coefficients and information-theoretic directional influence. This suggests a broader interpretive bridge: causal strength can sometimes be expressed either as a parametric connectivity magnitude or as an information-flow asymmetry, provided the modeling assumptions are restrictive enough (Wang et al., 2017).

4. Transient neural events, event propagation, and relative DCS

A central motivation for DCS is the analysis of brief, event-locked, non-stationary neural phenomena such as sharp-wave ripples. The transient-events paper emphasizes that DCS is designed for peri-event ensembles and that it performs better than TE in strongly synchronized or nearly deterministic settings. In simulations of synchronized oscillators, TE underestimates causal influence when synchrony is strong, whereas DCS stays high in the true direction and better reflects the coupling (Shao et al., 2022).

At the same time, that paper identifies a specific limitation: DCS is insensitive to deterministic perturbations in the source when applied to recurring spontaneous transient events. The reason is a shift-invariance argument. If the source event introduces a reproducible deterministic waveform but does not change the causal coupling coefficient itself, then both the observational and interventional distributions are translated together, and the DCS KL divergence remains essentially unchanged. In the paper’s phrasing, DCS measures connectivity, not event amplitude. This is precisely the regime in which neuroscientific interest often lies in event propagation rather than in a fixed connectivity parameter (Shao et al., 2022).

To address that limitation, the paper introduces relative Dynamic Causal Strength (rDCS) by replacing the source distribution at time $X \to Y$4 with a reference distribution from a baseline time $X \to Y$5. The effect is to compare the current event-driven source state against a baseline no-event state, making the measure sensitive to event-locked changes in the source. In simulations with a Morlet-wavelet perturbation of the source innovations, TE remains approximately constant, DCS remains approximately constant, and only rDCS shows the rhythmic event-related modulation (Shao et al., 2022).

The term “rDCS” is not completely uniform across the cited literature. TranCIT describes relative DCS as a normalized form of DCS used to express directional dominance across pairs, sessions, or conditions, for example

$X \to Y$6

By contrast, the transient-events paper defines rDCS through a baseline-reference intervention. This indicates that “relative DCS” is literature-specific rather than fully standardized: in one usage it denotes normalization across directions, and in another it denotes contrast against a reference peri-event state (Nouri et al., 30 Aug 2025, Shao et al., 2022).

5. Estimation pipelines and empirical use in neuroscience

TranCIT packages DCS into a transient causal interaction analysis pipeline. The sequence described in the paper is: data acquisition and preprocessing; event detection and alignment; event segmentation; pairwise causal analysis for each source-target pair; null-model or surrogate significance estimation; optional rDCS computation for relative directionality; and visualization as time-resolved causal trajectories, heatmaps, or directionality indices. The package implements Granger Causality, Transfer Entropy, DCS, and rDCS within the same Python toolbox, with the stated aim of making event-specific causal analysis accessible in the Python ecosystem (Nouri et al., 30 Aug 2025).

Its principal biological application is the hippocampal CA3–CA1 pathway during sharp-wave ripples. The paper states that this circuit is known to support coordinated population bursts and directional communication from CA3 to CA1 during SWRs, and uses it as a testbed for transient causal analysis. DCS is applied to estimate dynamic causal influence across SWR events, with the stated aim of determining whether CA3 drives CA1, when that drive occurs within the ripple, and how strong it is relative to the reverse direction. The reported outcome is that DCS and rDCS reveal the dominant CA3 $X \to Y$7 CA1 direction during the ripple and do so more clearly than traditional methods in short, high-synchrony regimes (Nouri et al., 30 Aug 2025).

A related but methodologically distinct empirical line appears in the DDCM–DI paper. There, fMRI data from resting-state and stimulus-based visual conditions are analyzed over the regions V1, PPA, and SMC. The reported result is that under visual stimulation $X \to Y$8 increases substantially relative to $X \to Y$9, indicating stronger influence from V1 to PPA, with similar behavior for V1 and SMC, and the paper states that the results are consistent with previous DI-based findings. Although this work does not compute DCS as such, it is directly relevant to causal-strength-type inference in neuroscience because it links model coefficients and directional information measures under explicit assumptions (Wang et al., 2017).

6. Assumptions, interpretation, and terminological scope

DCS is only as strong as its causal and statistical assumptions. TranCIT lists, at minimum, SCM validity, temporal ordering, local causal sufficiency or controlled confounding, intervention semantics, adequate sampling, and event consistency. The transient-events paper likewise assumes an SCM with independent exogenous noises, a time-inhomogeneous VAR representation, a soft intervention that replaces the source past by an independent copy, a valid marginal distribution for the source, and i.i.d. event-ensemble samples at each peri-event time. Both sources therefore treat DCS as an operational approximation of causal influence rather than a guarantee of absolute mechanistic truth (Nouri et al., 30 Aug 2025, Shao et al., 2022).

The broader literature on dynamic causal discovery provides a nearby but distinct perspective. DyCausal learns a time-varying weighted adjacency matrix trajectory $t$0, where the magnitude of each time-indexed edge weight is interpreted as evolving causal influence. The paper explicitly describes these weights as the closest analogue of DCS in its setting. This suggests a convergence between intervention-based DCS and dynamic graph-learning approaches: one emphasizes interventional divergence, the other emphasizes time-resolved edge coefficients, but both aim to recover trajectories of causal influence rather than a single static graph (Yang et al., 26 Feb 2026).

The acronym “DCS” is also polysemous in arXiv usage. Outside Dynamic Causal Strength, it denotes the Decentralization–Consensus–Scale triangle in blockchain theory (Slepak et al., 2018), dynamical Chern-Simons gravity in black-hole perturbation theory (Srivastava et al., 2021), and Deterministic Causal Structure in asynchronous multi-agent systems (Ren et al., 7 Oct 2025). For that reason, the full expression “Dynamic Causal Strength” is often necessary to avoid conflating unrelated literatures.

In its most specific and technically developed sense, Dynamic Causal Strength denotes an SCM-based, time-resolved, interventional measure that quantifies how much a target distribution changes when a source-to-target causal input is cut or perturbed. Its main contribution is to separate mechanistic causal influence from predictive dependence in transient neural data; its main caveat is that the precise operational form, especially for “relative” variants, depends on the modeling assumptions and the literature-specific formulation (Shao et al., 2022, Nouri et al., 30 Aug 2025).