收敛交叉映射（convergent cross mapping，CCM）滥觞、2012年Science论文方法部分：Detecting Causality in Complex Ecosystems

Dynamic causation and CCM

第1段

GC applies if the world is purely stochastic. However, to the extent that it is deterministic and dynamics are not entirely random, there will be an underlying manifold governing the dynamics (representing coherent trajectories as opposed to a random tangle). 如果世界是纯随机的，则GC适用。然而，在一定程度上，它是确定性的，并且动力学不完全是随机的，因此将有一个控制动力学的基本流形（代表连贯的轨迹，而不是随机纠缠）。

第2段

In dynamical systems theory, time-series variables (say, X and Y) are causally linked if they are from the same dynamic system (4, 19, 20)—that is, they share a common attractor manifold M (movies S1 to S3 illustrate this idea). This means that each variable can identify the state of the other (3, 19, 20, 24, 25) (e.g., information about past prey populations can be recovered from the predator time series, and vice versa). Additionally, when one variable X is a stochastic environmental driver of a population variable Y, information about the states of X can be recovered from Y, but not vice versa. For example, fish time series can be used to estimate weather, but not conversely. This runs counter to Granger’s intuitive scheme (see explanation in box S1). 在动力系统理论中，如果时间序列变量（例如，X和Y）来自同一个动力系统（4、19、20），则它们是因果关联的——也就是说，它们共享一个共同的吸引子流形M（movies S1~S3说明了这一想法）。这意味着每个变量可以识别另一个变量的状态（3、19、20、24、25）（例如，关于过去猎物种群的信息可以从捕食者时间序列中恢复，反之亦然）。此外，当一个变量X是人口变量Y的随机环境驱动因素时，有关X状态的信息可以从Y中恢复，但反之亦然。例如，fish时间序列可以用来估计天气，但不能反过来。这与格兰杰的直觉方案背道而驰（见方框S1中的解释）。

第3段

Our alternative approach, convergent cross mapping (CCM), tests for causation by measuring the extent to which the historical record of Y values can reliably estimate states of X. This happens only if X is causally influencing Y. In more detail, CCM looks for the signature of X in Y’s time series by seeing whether there is a correspondence between the “library” of points in the attractor manifold built from Y, M_Y, and points in the X manifold, M_X, where these two manifolds are constructed from lagged coordinates of the time-series variables Y and X, respectively (3, 19, 24) (movies S1 and S2). 我们的方法，收敛交叉映射（CCM），通过测量Y值的历史记录能够可靠估计X状态的程度来测试因果关系。只有当X对Y产生因果影响时才会发生这种情况。更详细地说，CCM通过查看由Y、M_Y构建的吸引子流形中的点的“库”与X流形中的点M_X之间是否存在对应关系来寻找Y时间序列中X的特征，其中这两个流形分别由时间序列变量Y和X的滞后坐标构建（3、19、24）（movies S1和S2）。

第4段

Essentially, the idea is to see whether the time indices of nearby points on the Y manifold can be used to identify nearby points on M_X. If so, then one can use Y to estimate X and vice versa. This procedure is illustrated in Fig. 2 and movie S3, with full technical details including an algorithm in (26). 本质上，这个想法是看Y流形上邻近点的时间索引是否可以用于识别M_X上的邻近点。如果是这样，那么可以使用Y来估计X，反之亦然。该过程在Fig. 2和movie S3中示出，其全部技术细节包括（26）中的算法。

第5段

Note that CCM is related to the general notion of cross prediction (3, 25) but with important differences. First, CCM estimates “states” across variables and does not forecast how the system “evolves” on the manifold. This eliminates possible information loss from chaotic dynamics (Lyapunov divergence) and accommodates nondynamic (i.e., random) variables. More important, CCM involves convergence, a key property that distinguishes causation from simple correlation. Convergence means that cross-mapped estimates improve in estimation skill with time-series length L (sample size used to construct a library) (Fig. 3A, fig. S2, and box S1). With more data, the trajectories defining the attractor fill in, resulting in closer nearest neighbors and declining estimation error (a higher correlation coefficient) as L increases (Fig. 2). Thus, CCM becomes a necessary condition for causation. Indeed, failure to account for convergence explains conflicting results reported in the literature with related methods (supplementary text and fig. S5). 注意，CCM与交叉预测（3，25）的一般概念有关，但有重要区别。首先，CCM跨变量估计“状态”，而不预测系统如何在流形上“进化”。这消除了混沌动力学（李雅普诺夫散度）可能造成的信息损失，并适应了非动力学（即随机）变量。更重要的是，CCM涉及收敛性，这是区分因果关系和简单相关性的关键属性。收敛意味着交叉映射估计在时间序列长度L（用于构建库的样本量）的估计技能上有所提高（Fig. 3A、fig. S2和box S1）。数据越多，定义吸引子的轨迹就会填充，导致最近邻越近，估计误差（更高的相关系数）随着L的增加而减小（Fig. 2）。因此，CCM成为因果关系的必要条件。事实上，未能解释收敛性解释了文献中报告的与相关方法相冲突的结果（补充文本和fig. S5）。

第6段

In practical applications, where shadow manifolds are low-dimensional approximations of the true system, convergence will be limited by observational error, process noise, and time-series length L. Thus, with limited or noisy field data, CCM is demonstrated by predictability that increases with L (fig. S3). See (26) for a discussion of data requirements. 在实际应用中，如果阴影流形是真实系统的低维近似值，则收敛将受到观测误差、过程噪声和时间序列长度L的限制。因此，对于有限或有噪声的现场数据，CCM通过随L增加的可预测性来证明（fig. S3）。有关数据要求的讨论，请参见（26）。