Common solar wind drivers behind magnetic storm–magnetospheric substorm dependency


The dynamical relationship between magnetic storms and magnetospheric substorms is one of the most controversial issues of contemporary space research. Here, we address this issue through a causal inference approach to two corresponding indices in conjunction with several relevant solar wind variables. We find that the vertical component of the interplanetary magnetic field is the strongest and common driver of both storms and substorms. Further, our results suggest, at least based on the analyzed indices, that there is no statistical evidence for a direct or indirect dependency between substorms and storms and their statistical association can be explained by the common solar drivers. Given the powerful statistical tests we performed (by simultaneously taking into account time series of indices and solar wind variables), a physical mechanism through which substorms directly or indirectly drive storms or vice versa is, therefore, unlikely.


The identification of spurious associations and potentially causal relationships is key to an improved process-based understanding of various geoscientific processes. Specifically in magnetospheric physics, the understanding of the relationship between magnetic storms and magnetospheric substorms as a part of the solar wind–magnetosphere system is of paramount importance for the development of numerical simulation models of the magnetosphere1. In particular, the existence and directionality of the storm – substorm interaction is one of the most controversial aspects of magnetospheric dynamics2. The original concept of storms being the cumulative result of successive substorms put forward by Akasofu in 19613 has been disputed in subsequent analyses2,4,5. While several model-based studies have shown a distinct impact of substorm injections on the storm-time ring current enhancement6,7,8, other studies have suggested that the ring current buildup could in principle be directly driven by the solar wind electric field9,10. In this case, magnetospheric substorms do not drive magnetic storms and the two phenomena are independent and share a common cause–the southward interplanetary magnetic field (IMF) driver11.

Recent work (see for instance the review by Balasis et al.12) points to a considerable importance of entropy-based measures for identifying and quantifying linear and nonlinear interdependencies between different geophysical variables, variability at different scales, and other characteristics. Time series analyses based on information-theoretic measures have been used to shed light on the storm-substorm interaction13 and the solar wind drivers of the outer radiation belt14 through the general perspective of quantifying information transfer, including linear and nonlinear mechanisms. In particular, DeMichelis et al.13 applied a bivariate transfer entropy15 (bivTE) analysis to the geomagnetic activity indices AL and SYM-H. SYM-H is the high-resolution (1-min) version of the hourly Disturbance storm-time (Dst) index, which is used as a proxy of magnetospheric ring current strength and, thus, as a measure of magnetic storm intensity. AL belongs to the set of the 1-min Auroral Electrojet indices (AE, AL, AU and AO) and is used to determine the onset of the substorm growth phase16. DeMichelis et al. suggested that information flow from AL to SYM-H dominates in the case of small geomagnetic disturbances, while the reverse situation is observed in presence of strong geomagnetic disturbances.

However, bivariate measures such as mutual information (MI) or bivTE do not allow to exclude the very frequent influence of other variables as common drivers, rendering MI and bivTE associations spurious. Multivariate extensions of TE, on the other hand, are severely limited because their estimators don’t work well in high dimensions17. In the present study, we contrast bivariate measures with a directional, multivariate information-theoretic causality measure based on low-dimensionally estimated graphical models17,18,19. This multivariate measure for the influence of a subprocess X of a system on another subprocess Y is called information transfer to Y (ITY) and allows for more powerful tests on the absence or potential presence of a causal relationship, which is crucial for developing a better “mechanistic” understanding of the governing processes.

Here, we investigate time series of various solar wind parameters including the IMF’s magnitude B and vertical component BZ, its velocity VSW and dynamic pressure Pdyn, as well as the AL and SYM-H indices. We focus on data from 2001, near a solar activity maximum. Our goal is to clarify whether substorm activity could causally drive storm dynamics or – on the contrary – whether solar wind variables can explain the statistical associations between storm and substorm activity.



The present study is focused on the year 2001, a year with very strong solar activity, and data on solar wind parameters as well as geomagnetic activity indices. As possible solar driving factors we include only those quantities with at least a measurable statistical dependency (mutual information) with either AL or SYM-H. These are B, BZ, VSW, and Pdyn. The original one-minutely data were aggregated to a 20-minute time resolution by averaging over non-overlapping 20-minute blocks. This resolution was selected based on iterative tests to obtain a compromise between resolving time lags and still keeping the computational load and multiple testing problems low. Additionally, DeMichelis et al.13 found, on average, a net information flow from AL to SYM-H attaining its maximum at a typical time delay of about 1 h which is well resolved with our chosen time resolution.

As solar wind time series inevitably contain missing values due to satellite failures, in the aggregation we masked samples for 20 min periods with more than 50% missing values. We also accounted for masked samples in the lagged analyses (up to τmax = 6 × 20 min) to avoid a selection bias. This leads to 18,384 non-masked 20-min samples instead of about 26,000 samples for the whole year. No further pre-processing was applied. Figure 1 shows the corresponding time series.

Figure 1

Time series of solar and magnetospheric variables for 2001. Time points with missing values in any of the variables are excluded from the analysis, taking lags into account. Clearly, there is strong solar and magnetospheric activity, in Supplementary Table S1 we classify storms into moderate, intense, and super-storm events.

Mutual information and bivariate transfer entropy analysis

This study aims to shed light on the possible existence of a driver-response relationship between storms and substorms, which is a reflection of the dynamic processes within the coupled solar wind-magnetosphere-ionosphere system. Because there has been accumulating evidence that the involved interrelations are of a nonlinear nature20,21,22,23,24,25,26 and very long data series are available, we employ a non-parametric (model-free) approach here. Information theory provides a genuine framework for the model-free study of couplings among time series. Here we invoke three information-theoretic measures with increasing power to detect spurious dependencies due to autocorrelation, common drivers or indirect relationships.

The first and simplest association measure applying information theory to time series is the lagged (cross-)mutual information27 given by

$${I}_{XY}^{{\rm{MI}}}(\tau )=I({X}_{t-\tau };{Y}_{t})=H({Y}_{t})-H({Y}_{t}|{X}_{t-\tau }),$$

using Shannon entropies \(H(X)=-\,\int \,p(x)\,\mathrm{ln}\,p(x)dx\) (correspondingly for conditional entropies) in units of nats with the natural logarithm as a base. For τ > 0, MI measures the information in the past of X that is contained in the present of Y. The weaknesses of MI as a measure of information transfer have been discussed early on, most notably by Schreiber15. A first step to arrive at a directional notion of information transfer is to exclude information from the past of Y. Implementing this idea, Schreiber introduced the transfer entropy (TE)15 between two variables, which is the information-theoretic analogue of Granger causality and can be defined in a lag-specific variant as

$${I}_{X\to Y}^{{\rm{bivTE}}}(\tau )=I({X}_{t-\tau };{Y}_{t}|{Y}_{t-1})$$

based on the conditional mutual information. To estimate all CMIs in this study, we use an advanced nearest-neighbor estimator28,29 that is most suitable for variables with a continuous range of values (details in Methods).

However, bivTE can yield spurious results if more than two processes are interacting: For the interaction example in Fig. 2(b) both the MI I(Xt−1;Yt) and the TE I(Xt−1;Yt|Yt−1) are larger than zero due to the common driver Z, even though no direct or indirect physical mechanism exists by which X drives Y or vice versa. The detailed time-resolved graph in Fig. 2(a) shows that, Xt−1 and Yt are not independent given only the past of Y or only the common driver Zt−2 as a condition. Rather, in order to unveil the spurious dependency, the CMI must be conditioned on both Yt−1 and Zt−2 to exclude all causal paths connecting Xt−1 and Yt (see ref.30 for a definition of causal paths).

Figure 2

Example of causal interactions in a three-variable process. (a) Time series graph30 which encodes the spatio-temporal dependencies. The set of parents \({{\mathscr{P}}}_{{Y}_{t}}\) (blue boxes) separates Yt from the past of the whole process \({{\bf{X}}}_{t}^{-}\backslash {{\mathscr{P}}}_{{Y}_{t}}\), which implies conditional independence (Markov property) and is used in the algorithm to estimate the graph17,19. (b) Process graph, which aggregates the information in the time series graph for better visualization (labels denote the lags, link and node colors denote the cross- and auto-coupling strength).

In Fig. 3 we investigate bivariate MI and bivTE lag functions of all considered solar variables with AL and SYM-H, including the interaction between these two. We restrict the maximum time delay to τmax = 6 × 20 min. For example, the panel BZ → AL shows the lag function I(BZ,tτ; ALt) of MI [Eq. (1), gray] and I(BZ,tτ; ALt|ALt−1) of bivTE excluding the past lag of AL [Eq. (2), black]. The multivariate ITY [Eq. (3), blue] is discussed in the next section. The solid line marks the significance threshold. All (C)MI values have been rescaled to the (partial) correlation scale via \(I\to \sqrt{1-{e}^{-2I}}\in [0,1]\)27 and rescaled values above 0.4 can, thus, be considered as moderate to strong. In the tables (Table 1 in main article and Supplementary Tables S2–S4), on the other hand, the CMI values are given in nats.

Figure 3

Lag functions of information-transfer measures. The lag functions were estimated with nearest-neighbor CMI estimation parameter k = 5028,29. For example, the panel BZ → AL shows the lag function \(I({B}_{Z,t-\tau };{{\rm{AL}}}_{t}|\ldots )\) of MI (Eq. (1), gray), bivTE excluding the past lag of AL (Eq. (2), black), and the multivariate ITY (Eq. (3), blue) conditioning out the influence also of other variables with the parents \({\mathscr{P}}\) given in Table 1. All (C)MI values have been rescaled to the (partial) correlation scale via \(I\to \sqrt{1-{e}^{-2I}}\in [0,1]\)27. For ITY, the solid line marks the significance threshold. MI and bivTE are clearly significant for a large range of lags. Confidence intervals (errorbars) are mostly smaller than the dots. MI and bivTE with their broad peaks clearly provide no precise information about relevant drivers and coupling delays. On the other hand, ITY features large values only at few selected lags. In Supplementary Fig. S1 we show that these results are robust for further method parameters and storm indices.

Table 1 Multivariate information-theoretic analysis of dependency between AL and SYM-H.

In Fig. 3, MI lag functions (grey) show large values for all possible driver variables. Here the peak of the MI lag function is often shifted compared to bivTE (black). Such an effect can arise from strong autocorrelations as studied in ref.31. Overall, the bivariate TE has sharper peaks than MI. BZ clearly is the strongest driver of both AL and SYM-H, and all other drivers are comparably weak (except for the auto-dependencies in panels AL → AL and SYM-H → SYM-H). The reason for this behavior is that some MI values are ‘inflated’, again, due to strong autocorrelations18, especially Vsw is strongly auto-dependent. This makes MI values and the peak of MI lag functions hard to interpret.

The interactions AL → SYM-H and SYM-H → AL have been studied in ref.13 where a relationship from substorms towards storms was found with a binning estimator of bivTE. Our results reproduce this finding with a nearest-neighbor estimator28,29. The other direction, from storms to substorms, is not very significant here. Note that values at lag τ = 0 min cannot be interpreted in a directional sense in our analysis.

Multivariate information-theoretic causality analysis

Mutual information and bivariate information-theoretic measures, such as MI and bivTE, cannot account for common drivers and indirect transitive relationships. As illustrated in Fig. 2(b), a common driver (Z) can lead to a spurious association, either linear or nonlinear, between X and Y. The complex multivariate causal interaction structure can be captured with the concept of a time series graph32,33 as shown in Fig. 2(a), originating from the theory of graphical models. As further defined in ref.30, each node in a time series graph represents a subprocess at a certain time. Past nodes at t′ < t have a link towards a subprocess at time t if and only if they are not independent conditionally on the past of the whole process. In this graph the parents \({{\mathscr{P}}}_{\cdot }\) of a variable are given by all nodes with an arrow towards it (blue boxes in Fig. 2(a)).

While these parents could be estimated by testing the CMI between each Xtτ and Yt conditional on all other lagged variables, this approach, similar to multivariate or conditional TE, does not work well due to its high dimensionality17 leading to weak statistical power and many false positives. In ref.19 an efficient algorithm for the estimation of the parents of a variable Y (further details in Methods) is detailed. In a second stage we use the estimated set of parents to measure the information transfer to Y (ITY)18 for all lagged variables Xtτ (including the parents)

$${I}_{X\to Y}^{{\rm{ITY}}}(\tau )=I({X}_{t-\tau };{Y}_{t}|{{\mathscr{P}}}_{{Y}_{t}}),$$

which will be zero if and only if Xtτ and Yt are independent conditionally on \({{\mathscr{P}}}_{{Y}_{t}}\). Unfortunately, no analytical results exist on the finite-sample distribution of the nearest-neighbor estimator under the null hypothesis of conditional independence. For significance testing, we use a block-shuffle surrogate test here following refs.34 and35 as described in Methods. The algorithm was run with maximum lag τmax = 6 × 20 min as before. We assess significance at the 95% level.

Table 1 shows iteration steps with the selected conditions and the conditional mutual information (CMI) values and significance of the AL → SYM-H and SYM-H → AL links in each step. The AL → SYM-H link becomes non-significant using the condition set (SYM-H(t − 1), BZ(t − 2), Pdyn(t − 1), Vsw(t − 2)). This implies that these solar drivers can explain the spurious link AL → SYM-H at a lag of 20 min. Note that this set is only a sufficient explanatory set and other drivers might also induce this spurious association. Also the much weaker link SYM-H → AL becomes non-significant after including few solar drivers (BZ, Vsw, B, at different lags).

The ITY estimates with these parents are shown in Fig. 3 (blue markers). ITY now accounts for autocorrelation in the driven variable (like bivTE), but additionally for the influence of the other parents as common drivers or indirect mediators36. Now the ITY lag functions are peaked and significant (markers above solid line) only at a few selected lags.

Figure 4 visualizes the significant drivers of AL and SYM-H in a process graph as in Fig. 2(b). Edges correspond to directional lagged links, and the labels indicate their lags. If more than one lag is significant, they are listed in the order of their strength. Both, the edge color and width, indicate the value at the lag with the largest ITY. The node color depicts the strength of the lag-1 auto-dependency for AL and SYM-H. Note that the weak ITY value in BZ → SYM-H is due to BZ occurring with two neighboring lags in the parents of SYM-H, which reduces the information transfer of either of them.

Figure 4

Graph based on significant ITY values at the 95% level in Fig. 3. Edges correspond to directional lagged links, and the labels indicate their lags. If more than one lag is significant, they are listed in the order of their strength. The edge color and width indicate the value at the lag with the largest ITY. The node color depicts the strength of the lag-1 auto-dependency for AL and SYM-H. Note that the weak ITY value in BZ → SYM-H is likely due to BZ occurring with two neighboring lags in the parents of SYM-H, which reduces the information transfer of either of them. In Supplementary Fig. S2 we show the robustness of these results using a different CMI estimation parameter and another substorm index.

In conjunction with some further robustness studies for another substorm index and other method parameters (Supplementary Figs. S1 and S2), our major results can be summarized as follows: The main drivers of substorms as measured by AL are BZ and VSW. These also drive storms as measured by SYM-H. Pdyn and especially B are less robustly related to both storms and substorms. Regarding time lags, the AL index responds to BZ at a lag ≈ 20–40 min, while the lags with the weaker other drivers are less robust (see Supplementary Figs. S1 and S2). The SYM-H index responds to BZ at a lag ≈ 40 min, to VSW at 40 min, to Pdyn at 20 min, and rather weakly with non-robust lags to B.

Regarding the previously found link AL → SYM-H13, we find that mainly BZ and to a lesser degree VSW and Pdyn are sufficient to explain this statistical association. These results are also verified by applying the same tools to an AE - SYM-H analysis and for other estimation parameters (see Supplementary Figs. S1 and S2 and Tables S2S4). Thus, we find that there is no direct or indirect transfer of information AL → SYM-H or SYM-H → AL.


DeMichelis et al.13 investigated the transfer of information between substorms and storms by means of a bivariate transfer entropy analysis of AL and SYM-H time series from 1981 (near solar maximum). They found a significant information flow from substorms to storms attaining its maximum with a typical time delay of about 1 h and suggested that the direction of information flow between substorms and storms depends on the global magnetospheric activity level. Our analysis goes beyond the study of ref.13 by utilizing a directional, multivariate information-theoretic causality measure that simultaneously takes into account solar wind variables and geomagnetic indices data, allowing for more powerful statistical tests on the absence or potential presence of a causal relationship between substorms and storms.

Our secondary finding that the main drivers of substorms (as measured by AL) and storms (as measured by SYM-H) are BZ and VSW is consistent with the fact that the energy transfer from the solar wind to the magnetosphere is proportional to BZ and VSW. Pdyn and especially B are less relevant for both storms and substorms. We conclude that these directed information transfers constitute robust interrelationships between solar wind parameters and dynamic processes in the magnetosphere. These findings confirm earlier studies on solar wind drivers and their storm and substorm manifestations, including variations of the indices SYM-H and AL (refs.37,38 and refs. therein).

Johnson et al.39 recently showed that the transfer of information from VSW or VBsouth (where VBsouth is VSW × southward IMF BZ) to Dst (similar to SYM-H but with a lower resolution) lasts more than 100 hours, which may correspond to the long time scale of the ring current decay (e.g., ref.40). We have also considered similar time scales when having analyzed VBsouth and Dst for another study focusing solely on storms41. Here, we analyze 1 minute-resolution indices and solar wind data in order to look at shorter time scales that cover lags up to 2 hours, focusing both on storms and substorms. The finding of IMF BZ and VSW as the drivers of storms are consistent with Johnson et al.39, which also used information theoretic tools (transfer entropy and cumulant based analysis) in their analysis.

Our most important finding is that our iterative causal discovery algorithm analysis suggests that mainly BZ, and to a lesser degree VSW and Pdyn are sufficient to explain the previously found link AL → SYM-H13. Thus, we find no statistical evidence for a link AL → SYM-H. We also find no link SYM-H → AL and these results are robust also for another substorm index (AE) and for other estimation parameters. The results by Iyemori and Rao42 supported the idea that the geomagnetic storms and substorms are independent processes; that is, the ring-current development is not the result of the frequent occurrence of substorms, but that of enhanced convection caused by the large southward IMF. Although some later studies43,44, based on in-situ observations, have shown that the contribution of ion injections to the ring current energy gain is substantial, our results do not favor the role of substorms in the enhancement of the storm-time ring current through accumulative ion injections during consecutive substorms, in agreement with the earlier studies by Iyemori and Rao42. A possible reason for the absence of information transfer from AL → SYM-H might be that not all ion injections to the storm-time ring current are reflected in the AL variations. A recent study45 showed that small-scale injections are not captured by AL. Another study46 showed that low- and high-energy protons vary in quite different ways on storm-time timescales and accordingly suggested that the relation between ion injections and ring current growth may be more complicated than previously perceived. In summary, it is possible that substorms are required for the particle injection to the ring current [e.g. seen in Energetic Neutral Atom imaging] but not sufficient (since non-storm substorms appear to lead to no intensification of the ring current) and strong convection is also required. This remains a debate.

Before concluding, let us discuss the methodological limitations pertaining to such a statistical causality analysis. The presence of significant links in our analysis can be called causal only with respect to the included set of variables. Non-observed variables can still be the cause of a link here and the obtained links should, therefore, serve more as an hypothesis for further studies that include more possible explanatory variables. From a theoretical standpoint, a more robust finding is that of the absence of a link: if there exists no statistical evidence for a dependency between two variables, a physical mechanism between the two is unlikely. Hence, the non-significance of direct or indirect dependencies between the commonly used AL and SYM-H indices leads us to the conclusion that there exists no physical mechanism by which perturbations in substorms are transported to storms or vice versa. However, from a practical standpoint in the general context of the limitations associated with every statistical information quantity, below we summarize possible deficiencies and/or weaknesses that may accompany the application of ITY, even though we consider them rather insignificant or of low probability to occur for the present study.

Firstly, the information measure might not capture the dependency. We should note that our information-theoretic approach allows to take into account almost any type of nonlinear relationship, both in excluding it as a common driver, and also in detecting it. This is in contrast to linear correlation or linear Granger causality studies. The price for this “generality” is lower statistical power: For a particular dependency, the more general CMI will have less power compared to a measure that is optimized for this type of dependency, for example correlation for linear dependencies. Weaker power means that weaker dependencies might not be detected for small sample sizes, especially for high-dimensional conditions30. Our method is designed to avoid high-dimensionality by an iterative approach (especially compared to multivariate TE) and has demonstrated high power in numerical experiments30. Additionally, here we have a very large sample size, leading us to the conclusion that if there is a dependency, it must be very weak. Also, our major finding is robust when using other estimation parameters.

Secondly, we analyzed the whole year 2001 to obtain a sufficiently large sample size. Possibly, a causal relationship is present only during shorter periods and absent in other periods, which would be difficult to assess given too short sample sizes and the length of characteristic time scales of the processes.

Thirdly, the physical mechanism might be present mostly during the missing values excluded in the analysis. If satellite failures are indeed strongly related with the hypothesized mechanism, this would imply a non-avoidable selection bias in our analysis.

Lastly, the indices have serious limitations as to their ability of monitoring a particular current system: (1) They are scalars and may be insufficient to deduce a 2D current system; (2) The indices are derived from a very limited number of stations and as such are subject to a number of artifacts and limitations; (3) The ground perturbations are due to all currents – near and far. In summary, AL is a limited measure of the 2D westward electrojet distribution (intensity, structure and dynamics) and SYM-H is a limited measure of the ring current (intensity, structure and dynamics)47. Here we tested two kinds of indices for substorms (AL and AE) and got robust results.

These limitations (generality-power trade-off, missing values, proxy data quality) apply to any statistical coupling analysis. The main shortcoming of previous approaches based on bivariate measures is that these did not take into account possible common drivers, hence weakening a possible causal interpretation. Multivariate approaches have a stronger causal interpretation at the cost of weaker detection power due to higher dimensionality, which our method alleviates as much as possible. In light of these qualifications, we conclude that a direct or indirect physical mechanism by which substorms drive storms or vice versa is unlikely.


There has been only one study so far that utilized information theory tools to study the storm-substorm relation. De Michelis et al.13 used the bivariate measures of delayed mutual information and transfer entropy to analyze SYM-H and AL indices from a year near solar maximum (1981). Their findings suggested that information flow from AL to SYM-H dominates in the case of weak geomagnetic disturbances, while the reverse situation is observed in the presence of strong geomagnetic disturbances. The present study goes beyond the analysis performed by De Michelis et al.13 by contrasting bivariate measures with a directional, multivariate information-theoretic causality measure based on low-dimensionally estimated graphical models. This multivariate measure is called information transfer to Y and allows for more powerful tests on the absence or potential presence of a causal relationship, which is crucial for developing a better “mechanistic” understanding of the governing processes. Thus, we are able to simultaneously handle SYM-H and AL indices along with the magnetospheric activity solar wind variables including the IMF’s magnitude B and vertical component BZ, its velocity VSW and dynamic pressure Pdyn using an information measure technique. This is the first time, to our knowledge, that the variations of the various parameters describing the input and output of the solar wind – magnetosphere system are treated all together by a causality measure that is capable to identify information flow between all these parameters. We conclude on non-significant direct or indirect dependencies between AL and SYM-H indices, and therefore, between substorms and storms, which means that the previously applied bivariate measures were not able to accurately depict or resolve the interdependencies between the system’s parameters.

Additionally, a secondary conclusion of our study is that we are able to confirm earlier results about the solar wind drivers of storms and substorms (i.e., BZ) from a different point of view, utilizing the modern and versatile toolbox of information theory. We have achieved this, for the first time, by considering the solar wind-magnetosphere system as a whole and applying a multivariate information-theoretic approach able to simultaneously handle the system’s input (solar wind drivers) and output (magnetospheric activity indicators) in contrast to several previous important but distinct studies, where the bulk of information on the solar wind driver of the magnetosphere has been accumulated [e.g. refs.37,38]. This demonstrates the great potential that the application of information theory may have to treat space physics problems, where vast amounts of related datasets are continuously accumulated either from spaceborne or ground-based measurements. Moreover, in the light of our findings, the application of the multivariate causality measure of ITY was able to explain the spurious link AL → SYM-H found previously by bivariate causality measures13 simply by the variations of the solar wind drivers.

The results of this study contribute to the ongoing debate of the storm-substorm relationship and to the debate of plasma injection to the inner magnetosphere. For example, Angelopoulos et al.48 concluded that bursty-bulk flows (BBFs) are sufficient to account for all the energy deposition in the ionosphere and inner magnetosphere. However, Ohtani et al.49 reported that fast plasma sheet flows do not reach the geosynchronous orbit or lead to dipolarization. The results of the present study offer an interesting possibility that substorm led injections or BBFs, in general, may not travel all the way to the inner magnetosphere. However, substorm BBFs accompanied by strong convection (VBsouth) may penetrate the inner magnetosphere and contribute to the ring current.

Our analysis demonstrates the great potential of combining a causal discovery algorithm with a multivariate and lag-specific extension of transfer entropy for tackling contemporary research questions in magnetospheric physics, such as the storm-substorm relationship, which is one of the most controversial topics of magnetospheric dynamics and solar-terrestrial coupling. Further analyses using a causal pathway-analysis36,50 can shed light on the interaction mechanism among the solar drivers and the magnetosphere. The obtained causal drivers, on the other hand, can also be relevant for optimal prediction schemes51. We expect that our results will contribute to a better understanding of the dynamic processes related to the coupled solar wind - magnetosphere - ionosphere system by fostering a paradigm shift in our perception of the storm-substorm relationship. They may also have a direct impact on magnetosphere modeling and, consequently, space weather forecasting efforts.


The algorithm in ref.19 for the estimation of the parents of a variable Y uses the idea to successively test for conditional independence between Yt and each possible past driver (including the past of Y) conditioned on iteratively more conditions. Thereby, the condition dimension stays as low as possible in every iteration step which helps to alleviate high dimensionality in estimating CMIs. Here we test only the most relevant set of conditions with the highest CMIs in the previous step. The algorithm then is as follows: We first initialize the preliminary parents \({\mathscr{P}}({Y}_{t})=({{\bf{X}}}_{t-1},{{\bf{X}}}_{t-2},\ldots ,{{\bf{X}}}_{t-{\tau }_{{\rm{\max }}}})\) containing the past of all variables (including Y). Starting with p = 0, we iteratively increase p → p + 1 in an outer loop and, in an inner loop, test for all variables \({X}_{t-\tau }^{i}\) from \({\mathscr{P}}({Y}_{t})\) whether

$$I({X}_{t-\tau }^{i};{Y}_{t}|{{\mathscr{P}}}^{(p)}({Y}_{t}))=0$$

where \({{\mathscr{P}}}^{(p)}({Y}_{t})\) are the p strongest parents among \({\mathscr{P}}({Y}_{t})\backslash \{{X}_{t-\tau }^{i}\}\) according to their CMI. If the CMI is zero at some significance level α using the test described below, we remove a link from \({\mathscr{P}}({Y}_{t})\) at the end of each p-iteration. The algorithm converges if no larger conditioning sets can be tested. We sort \({\mathscr{P}}({Y}_{t})\) after every iteration according to the CMI values.

We use an advanced nearest-neighbor estimator28,29 of CMI that is most suitable for variables with a continuous range of values. This estimator has as a parameter the number of nearest-neighbors k which determines the size of hyper-cubes around each (high-dimensional) sample point and, therefore, can be viewed as a density smoothing parameter (although it is data-adaptive unlike fixed-bandwidth estimators). For large k, the underlying dependencies are strongly smoothed and may not resolve nonlinearities. We tested different values of k to verify the robustness of our results. Larger k have larger bias and are more computationally expensive, but have smaller variance. Note that the estimated CMI values can be slightly negative while CMI is a non-negative quantity. In Figs. 3 and 4 and Supplementary Figs. S1 and S2 the (C)MI values have been rescaled to the (partial) correlation scale via \(I\to \sqrt{1-{e}^{-2I}}\in [0,1]\)27. In the tables, on the other hand, the CMI values are given in nats.

For significance testing, either a fixed threshold or shuffle surrogates are the only choice here. Surrogate tests are especially helpful for proper significance tests because they adapt to the bias for higher-dimensional CMIs. In ref.17 a shuffle test has been used, but for strongly autocorrelated time series, as in the present case, this test is too weak. Therefore, we use a block-shuffle surrogate test here following refs.34 and35. An ensemble of M values of \(I({X}_{t-\tau }^{\ast };{Y}_{t}|Z)\) is generated where \({X}_{t-\tau }^{\ast }\) is a block-shuffled sample of Xtτ, i.e., with blocks of the original time series permuted. As an optimal block-length we use the approach described in refs.34 and35 for non-overlapping blocks. The optimal block-length (Eq. (6) in ref.35) involves the decay rate of the envelope of the autocorrelation function γ(τ). The latter is estimated up to a maximum delay of 5% of the (non-masked) samples and the envelope was estimated using the Hilbert transform. Then a function τ is fitted to the envelope with constant C to obtain the decay rate ϕ. Finally, the CMI values are sorted and a p-value is obtained as the fraction of surrogates with CMI greater or equal than the estimated CMI value. We use an ensemble of 200 surrogates. Confidence intervals (errorbars in figures) were estimated using bootstrap resampling involving only estimated nearest-neighbor statistics with 200 samples. The block-shuffle approach is only an approximation to obtain the true null distribution.

Software Availability

Software is available online under


  1. 1.

    Pulkkinen, T. I. et al. Solar wind - magnetosphere coupling: A review of recent results. J. Atmos. and Solar Terr. Phys. 69, 256–264 (2007).

    ADS  Article  Google Scholar 

  2. 2.

    Sharma, A. S. et al. The Storm-Substorm Relationship: Current Understanding and Outlook, in Disturbances in Geospace: The Storm-Substorm Relationship (American Geophysical Union, Washington DC, 2003).

    Google Scholar 

  3. 3.

    Akasofu, S., Cain, J. C. & Chapman, S. The Magnetic Field of a Model Radiation Belt, Numerically Computed. Journal of Geophysical Research 66, 4013–4026 (1961).

    ADS  Article  Google Scholar 

  4. 4.

    Gonzalez, W. D., Joselyn, J. A. & Kamide, Y. What is a geomagnetic storm? Journal of Geophysical Research: Space Physics 99, 5771–5792 (1994).

    Article  Google Scholar 

  5. 5.

    Kamide, Y. et al. Current understanding of magnetic storms: Storm-substorm relationships. Journal of Geophysical Research: Space Physics 103, 17705–17728 (1998).

    CAS  Article  Google Scholar 

  6. 6.

    Fok, M.-C., Moore, T. E. & Delcourt, D. C. Modeling of inner plasma sheet and ring current during substorms. Journal of Geophysical Research 104, 14557–14570 (1999).

    ADS  Article  Google Scholar 

  7. 7.

    Daglis, I. A. & Kamide, Y. The role of substorms in storm-time particle acceleration. In Kamide, Y. & Gurbax, S. (eds) Disturbances in Geospace: The Storm-Substorm Relationship (American Geophysical Union, Washington, DC, 2004).

    Google Scholar 

  8. 8.

    Ganushkina, N. Y., Pulkkinen, T. I. & Fritz, T. Role of substorm-associated impulsive electric fields in the ring current development during storms. Annales Geophysicae 23, 579–591 (2005).

    ADS  CAS  Article  Google Scholar 

  9. 9.

    Kamide, Y. Is substorm occurrence a necessary condition for a magnetic storm? Journal of Geomagnetism and Geoelectricity 44, 109–117 (1992).

    ADS  Article  Google Scholar 

  10. 10.

    McPherron, R. L., Baker, D. N., Bargatze, L. F., Clauer, C. R. & Holzer, R. E. IMF control of geomagnetic activity. Advances in Space Research 8, 71–86 (1988).

    ADS  Article  Google Scholar 

  11. 11.

    Daglis, I. A. et al. Intense space storms: Critical issues and open disputes. Journal of Geophysical Research: Space Physics 108 (2003).

  12. 12.

    Balasis, G. et al. Statistical Mechanics and Information-Theoretic Perspectives on Complexity in the Earth System. Entropy 15, 4844–4888 (2013).

    ADS  Article  Google Scholar 

  13. 13.

    De Michelis, P., Consolini, G., Materassi, M. & Tozzi, R. An information theory approach to the storm-substorm relationship. Journal of Geophysical Research: Space Physics 116 (2011).

  14. 14.

    Wing, S., Johnson, J. R., Camporeale, E. & Reeves, G. D. Information theoretical approach to discovering solar wind drivers of the outer radiation belt. Journal of Geophysical Research: Space Physics 121 (2016).

    ADS  Google Scholar 

  15. 15.

    Schreiber, T. Measuring information transfer. Physical Review Letters 85, 461–464 (2000).

    ADS  CAS  Article  Google Scholar 

  16. 16.

    Gjerloev, J. W., Hoffman, R. A., Friel, M. M., Frank, L. A. & Sigwarth, J. B. Substorm behavior of the auroral electrojet indices. Annales Geophysicae 22, 2135–2149 (2004).

    ADS  Article  Google Scholar 

  17. 17.

    Runge, J., Heitzig, J., Petoukhov, V. & Kurths, J. Escaping the Curse of Dimensionality in Estimating Multivariate Transfer Entropy. Physical Review Letters 108, 258701 (2012).

    ADS  Article  Google Scholar 

  18. 18.

    Runge, J., Heitzig, J., Marwan, N. & Kurths, J. Quantifying causal coupling strength: A lag-specific measure for multivariate time series related to transfer entropy. Physical Review E 86, 061121 (2012).

    ADS  Article  Google Scholar 

  19. 19.

    Runge, J., Nowack, P., Kretschmer, M., Flaxman, S. & Sejdinovic, D. Detecting causal associations in large nonlinear time series datasets. arXiv:1702.07007v2 1–46 (2018).

  20. 20.

    Bargatze, L. F., Baker, D. N., Hones, E. W. & Mcpherron, R. L. Magnetospheric impulse response for many levels of geomagnetic activity. Journal of 90, 6387–6394 (1985).

    Google Scholar 

  21. 21.

    Tsurutani, B. T., Goldstein, B. E., Sugiura, M., Iyemori, T. & Gonzalez, W. D. The nonlinear response of AE to the IMF Bs driver - A spectral break at 5 hours. Geophysical Research Letters 17, 279 (1990).

    ADS  Article  Google Scholar 

  22. 22.

    Consolini, G., Marcucci, M. & Candidi, M. Multifractal Structure of Auroral Electrojet Index Data. Physical Review Letters 76, 4082–4085 (1996).

    ADS  CAS  Article  Google Scholar 

  23. 23.

    Klimas, A. J., Vassiliadis, D., Baker, D. N. & Roberts, D. A. The organized nonlinear dynamics of the magnetosphere. Journal of Geophysical Research: Space Physics 101, 13089 (1996).

    Article  Google Scholar 

  24. 24.

    Vassiliadis, D., Klimas, A., Baker, D. N. & Roberts, D. The nonlinearity of models of the vBsouth-AL coupling NN, Xn + l F (Xn, Un). Journal of Geophysical Research: Space Physics 101, 779–787 (1996).

    Article  Google Scholar 

  25. 25.

    Hnat, B., Chapman, S. & Rowlands, G. Scaling of solar wind and the AU, AL and AE indices as seen by WIND. Geophysical Research Letters 29, 35–1 (2002).

    Google Scholar 

  26. 26.

    Balasis, G. et al. Investigating dynamical complexity in the magnetosphere using various entropy measures. Journal of Geophysical Research: Space Physics 114 (2009).

    Article  Google Scholar 

  27. 27.

    Cover, T. M. & Thomas, J. A. Elements of Information Theory (John Wiley & Sons, Hoboken, 2006).

  28. 28.

    Kraskov, A., Stögbauer, H. & Grassberger, P. Estimating mutual information. Physical Review E 69, 066138 (2004).

    ADS  MathSciNet  Article  Google Scholar 

  29. 29.

    Frenzel, S. & Pompe, B. Partial Mutual Information for Coupling Analysis of Multivariate Time Series. Physical Review Letters 99, 204101 (2007).

    ADS  Article  Google Scholar 

  30. 30.

    Runge, J. Causal network reconstruction from time series: From theoretical assumptions to practical estimation. Chaos An Interdiscip. J. Nonlinear Sci. 28, 075310 (2018).

    CAS  MATH  Google Scholar 

  31. 31.

    Runge, J., Petoukhov, V. & Kurths, J. Quantifying the Strength and Delay of Climatic Interactions: The Ambiguities of Cross Correlation and a Novel Measure Based on Graphical Models. Journal of Climate 27, 720–739 (2014).

    ADS  Article  Google Scholar 

  32. 32.

    Dahlhaus, R. Graphical interaction models for multivariate time series. Metrika 51, 157–172 (2000).

    MathSciNet  Article  Google Scholar 

  33. 33.

    Eichler, M. Graphical modelling of multivariate time series. Probability Theory and Related Fields 153, 233–268 (2012).

    MathSciNet  Article  Google Scholar 

  34. 34.

    Peifer, M. et al. On Studentising and Blocklength Selection for the Bootstrap on Time Series. Biometrical Journal 47, 346–357 (2005).

    MathSciNet  CAS  Article  Google Scholar 

  35. 35.

    Mader, M., Mader, W., Sommerlade, L., Timmer, J. & Schelter, B. Block-bootstrapping for noisy data. Journal of Neuroscience Methods 219, 285–91 (2013).

    Article  Google Scholar 

  36. 36.

    Runge, J. Quantifying information transfer and mediation along causal pathways in complex systems. Phys. Rev. E 92, 062829 (2015).

    ADS  MathSciNet  Article  Google Scholar 

  37. 37.

    Vassiliadis, D. Systems theory for geospace plasma dynamics. Reviews of Geophysics 44, RG2002 (2006).

    ADS  Article  Google Scholar 

  38. 38.

    Tsyganenko, N. A. Data-based modelling of the Earth’s dynamic magnetosphere: A review. Annales Geophysicae 31, 1745–1772 (2013).

    ADS  Article  Google Scholar 

  39. 39.

    Johnson, J. R., Wing, S. & Camporeale, E. Transfer entropy and cumulant-based cost as measures of nonlinear causal relationships in space plasmas: applications to dst. Annales Geophysicae 36, 945–952 (2018).

    ADS  Article  Google Scholar 

  40. 40.

    Kozyra, J. et al. Multistep dst development and ring current composition changes during the 4–6 june 1991 magnetic storm. Journal of Geophysical Research: Space Physics 107, SMP–33 (2002).

    Article  Google Scholar 

  41. 41.

    Balasis, G. et al. From pre-storm activity to magnetic storms: a transition described in terms of fractal dynamics. Annales Geophysicae 24, 3557–3567 (2006).

    ADS  Article  Google Scholar 

  42. 42.

    Iyemori, T. & Rao, D. R. Decay of the Dst field of geomagnetic disturbance after substorm onset and its implication to storm-substorm relation. Annales Geophysicae 14, 608–618 (1996).

    ADS  Article  Google Scholar 

  43. 43.

    Daglis, I. A. et al. “Fine structure” of the storm-substorm relationship: Ion injections during Dst decrease. Advances in Space Research 25, 2369–2372 (2000).

    ADS  CAS  Article  Google Scholar 

  44. 44.

    Gkioulidou, M. et al. The role of small-scale ion injections in the buildup of Earth’s ring current pressure: Van Allen Probes observations of the 17 March 2013 storm. Journal of Geophysical Research: Space Physics 119, 7327–7342 (2014).

    ADS  Google Scholar 

  45. 45.

    Gkioulidou, M. et al. Spatial structure and temporal evolution of energetic particle injections in the inner magnetosphere during the 14 July 2013 substorm event. Journal of Geophysical Research: Space Physics 120, 1924–1938 (2015).

    ADS  Google Scholar 

  46. 46.

    Gkioulidou, M., Ukhorskiy, A., Mitchell, D. G. & Lanzerotti, L. J. Storm-time dynamics of ring current protons: Implications for the long-term energy budget in the inner magnetosphere. Geophysical Research Letters 43, 4736–4744 (2016).

    ADS  Article  Google Scholar 

  47. 47.

    Newell, P. & Gjerloev, J. Supermag-based partial ring current indices. Journal of Geophysical Research: Space Physics 117 (2012).

    Article  Google Scholar 

  48. 48.

    Angelopoulos, V. et al. Multipoint analysis of a bursty bulk flow event on april 11, 1985. Journal of Geophysical Research: Space Physics 101, 4967–4989 (1996).

    Article  Google Scholar 

  49. 49.

    Ohtani, S., Singer, H. & Mukai, T. Effects of the fast plasma sheet flow on the geosynchronous magnetic configuration: Geotail and goes coordinated study. Journal of Geophysical Research: Space Physics 111 (2006).

  50. 50.

    Runge, J. et al. Identifying causal gateways and mediators in complex spatio-temporal systems. Nature Communications 6, 8502 (2015).

    CAS  Article  Google Scholar 

  51. 51.

    Runge, J., Donner, R. V. & Kurths, J. Optimal model-free prediction from multivariate time series. Phys. Rev. E 91, 052909 (2015).

    ADS  Article  Google Scholar 

Download references


We acknowledge the following sources of data: the Kyoto World Data Center (WDC) for Geomagnetism and the observatories that produce and make AE and SYM-H indices available at as well as NASA’s Space Physics Data Facility (SPDF) that hosts the Coordinated Data Analysis Web (CDAWeb) services along with the science teams that provided interplanetary data through J.R. received funding by the James S. McDonnell Foundation. R.V.D. acknowledges financial support via the BMBF-funded Young Investigators Group “Complex Systems Approaches to Understanding Causes and Consequences of Past, Present and Future Climate Change” (CoSy-CC2, grant no. 01LN1306A). This work has been financially supported by the joint Greek–German IKYDA 2013 project “Transdisciplinary assessment of dynamical complexity in magnetosphere and climate: A unified description of the nonlinear dynamics across extreme events” funded by IKY and DAAD. The authors thank Ciaron Linstead for help with high-performance computing and gratefully acknowledge the European Regional Development Fund (ERDF), the German Federal Ministry of Education and Research (BMBF) and the Land Brandenburg for supporting this project by providing resources on the high performance computer system at the Potsdam Institute for Climate Impact Research.

Author information




G.B., J.R., R.V.D. and I.A.D. designed the study, C.P. prepared the data, J.R. analyzed the data. All authors discussed the results and contributed to editing the manuscript.

Corresponding author

Correspondence to Jakob Runge.

Ethics declarations

Competing Interests

The authors declare no competing interests.

Additional information

Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Electronic supplementary material

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Runge, J., Balasis, G., Daglis, I.A. et al. Common solar wind drivers behind magnetic storm–magnetospheric substorm dependency. Sci Rep 8, 16987 (2018).

Download citation


  • Solar Wind Driver
  • Common Driver
  • Solar Wind-magnetosphere System
  • Transfer Entropy (TE)
  • Storm-time Ring Current

Further reading


By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.