Past-Future CCA & Deep Dynamic Modeling

• The paper introduces PFCCA by framing past-to-future inference with a shared latent state that captures how earlier observations constrain later ones.
• It employs both shared and view-specific latent factors with GRU-like gated transitions and state-dependent variances to model nonlinear dynamics.
• Empirical results on financial time series show that D²PCCA improves probabilistic fit (ELBO) over linear DPCCA while preserving the CCA-based latent decomposition.

Past-Future Canonical Correlation Analysis (PFCCA) denotes a dynamic extension of CCA-style multiview representation learning for sequential data, in which the central task is not merely to correlate two views at the same instant but to capture relationships between past and future segments. In the formulation associated with Deep Dynamic Probabilistic Canonical Correlation Analysis (D$^2$PCCA), PFCCA-style modeling seeks latent states that explain how earlier observations constrain later ones, using a shared latent factor for common dynamics and view-specific latent factors for idiosyncratic evolution. D$^2$PCCA is presented as a probabilistic, deep nonlinear dynamic generalization of PFCCA-style ideas: rather than directly maximizing correlation between past and future views in a deterministic way, it models the latent temporal process generatively and learns it with variational inference (Tang et al., 7 Feb 2025).

1. Problem setting and relation to CCA

Classical CCA considers two sets of variables, $x^1$ and $x^2$, and learns linear projections so that the transformed variables are maximally correlated. In that form, CCA is a static two-view relationship model. Probabilistic CCA (PCCA) replaces the deterministic correlation objective with a latent-variable generative model, typically involving a shared latent factor $\mathbf{z}^0$ that generates both views, and in a richer version additional private factors $\mathbf{z}^1,\mathbf{z}^2$ for the individual views (Tang et al., 7 Feb 2025).

PFCCA arises from the observation that sequential data are not static and that the key information often lies in how the system evolves over time. In that setting, the relevant target is a latent representation that explains relationships between past and future segments rather than only contemporaneous cross-view dependence. The stated motivation of dynamic CCA-style models and past-future methods is therefore to learn latent states that explain how earlier observations constrain later ones. This places PFCCA conceptually between static multiview correlation analysis and full latent-state time-series modeling.

The dynamic formulation discussed in D$^2$PCCA preserves the CCA intuition of shared and private structure while adding temporal transitions. There is a shared latent factor tracking common dynamics, view-specific latent factors capturing idiosyncratic evolution, and explicit temporal dependence. This suggests that PFCCA is best understood not as a departure from CCA, but as a temporal reorganization of the same multiview decomposition.

2. Latent-variable decomposition in the dynamic setting

For each time $t$, the dynamic model uses three latent variables: $z_t^0$, a shared latent state underlying both observed variables; $z_t^1$, a private latent state for $^2$0; and $^2$1, a private latent state for $^2$2. The observed variables are $^2$3 and $^2$4. This decomposition retains the core multiview logic of CCA while making the latent representation explicitly temporal (Tang et al., 7 Feb 2025).

The generative model is specified as

$^2$5

$^2$6

for $^2$7 and $^2$8.

This formulation implies three structural properties. First, each latent chain is first-order Markov. Second, each observed view depends jointly on the shared latent state and its own private latent state. Third, both the conditional means and the conditional variances are learned functions of the latent state. The paper also writes the model in compact form as

$^2$9

$x^1$0

Within the PFCCA perspective, the significance of this factorization is that shared temporal regularities and view-specific temporal effects are separated at the latent level. That separation is the principal reason the model remains recognizably CCA-like even after being extended to sequential data.

3. Deep nonlinear parameterization

D$x^1$1PCCA is a nonlinear extension of DPCCA. It keeps the same latent-variable structure but replaces linear Gaussian transitions and emissions with neural networks. The transition model is parameterized by a gated transition network: $x^1$2

$x^1$3

$x^1$4

$x^1$5

The paper states that this resembles a GRU-like mechanism: the gate $x^1$6 controls how much nonlinear update is used (Tang et al., 7 Feb 2025).

The emission model is likewise nonlinear: $x^1$7

$x^1$8

$x^1$9

Several consequences follow directly from these definitions. The state evolution is nonlinear, the view generation process is nonlinear, and the transition and emission variances are state-dependent. The paper characterizes this as the use of nonlinear dynamics together with heteroscedastic variances depending on latent state. In PFCCA terms, this replaces a purely correlation-based temporal matching criterion with a learned nonlinear generative mechanism for how shared and private latent factors evolve and generate future observations.

4. Past-future inference and the variational objective

Because the model is nonlinear, exact inference is not tractable, and D$x^2$0PCCA uses variational inference. The posterior factorization is given as

$x^2$1

The approximate posterior adopts the same structure: $x^2$2

$x^2$3

Here $x^2$4 is produced by a backward RNN over the observations and therefore encodes future context $x^2$5 (Tang et al., 7 Feb 2025).

The combiner used in the inference network is

$x^2$6

$x^2$7

$x^2$8

The training objective is the evidence lower bound,

$x^2$9

The paper describes its familiar decomposition as reconstruction or emission log-likelihood minus a KL divergence between the approximate posterior and the transition prior.

For PFCCA, the decisive feature is the dependence

$\mathbf{z}^0$0

which means that the latent state at time $\mathbf{z}^0$1 is inferred from both the previous latent state and future observations. The paper identifies this as exactly the kind of mechanism that captures past-to-future predictive structure. In that sense, PFCCA is not simply about fitting current correlations; it is about learning latent states that summarize how past context evolves into future outcomes.

5. Distinction from static PCCA and from naive dynamic reductions

The paper contrasts three levels of modeling. Standard CCA is static, linear, deterministic, and correlation-based. PCCA introduces latent variables and a probabilistic generative view, but remains static, two-view, and non-temporal. DPCCA and D$\mathbf{z}^0$2PCCA extend the probabilistic CCA logic into time: each latent state evolves over time, the same latent factor that links the two views at time $\mathbf{z}^0$3 is also temporally dependent on $\mathbf{z}^0$4, and each observation is generated from the current shared and private latent states (Tang et al., 7 Feb 2025).

A common simplification would be to treat the problem as an ordinary dynamic latent-variable model on concatenated observations. The paper explicitly argues against that reduction. It notes that a naive dynamic extension of static PCCA,

$\mathbf{z}^0$5

would become trivial if transitions and emissions remain spherical Gaussian, because the two views can effectively be collapsed into a single standard DMM.

D$\mathbf{z}^0$6PCCA avoids that collapse by maintaining shared and private latent structure, view-specific emissions, nonlinear dynamics, and heteroscedastic variances depending on latent state. This distinction is central to the PFCCA viewpoint. The model is not merely a DMM with a multivariate observation vector; it is a dynamic multiview model whose latent organization retains the decomposition associated with CCA.

6. Enhancements, empirical results, and scope

The model incorporates two training and inference enhancements. First, KL annealing is implemented through a $\mathbf{z}^0$7-weighted ELBO in which $\mathbf{z}^0$8 starts small and is gradually increased to $\mathbf{z}^0$9, with the stated purpose of avoiding posterior collapse. Second, the approximate posterior can be made more expressive with normalizing flows: $\mathbf{z}^1,\mathbf{z}^2$0 The paper states that it uses affine autoregressive flows and reports that they improve ELBO (Tang et al., 7 Feb 2025).

The experimental evaluation uses a financial time series dataset with daily closing prices, 5 sectors, 10 companies per sector, sequences of length $\mathbf{z}^1,\mathbf{z}^2$1, training on the first 453 days, and testing on the last 50. The compared models are DPCCA, D$\mathbf{z}^1,\mathbf{z}^2$2PCCA, D$\mathbf{z}^1,\mathbf{z}^2$3PCCA + KL annealing, D$\mathbf{z}^1,\mathbf{z}^2$4PCCA + IAF, and D$\mathbf{z}^1,\mathbf{z}^2$5PCCA + KL + IAF. The reported metrics are Test ELBO and Test RMSE.

Model	Test ELBO	Test RMSE
D$\mathbf{z}^1,\mathbf{z}^2$6PCCA + KL	130.07	0.0179
D$\mathbf{z}^1,\mathbf{z}^2$7PCCA + IAF	130.96	0.0183
D$\mathbf{z}^1,\mathbf{z}^2$8PCCA + KL + IAF	131.27	0.0184
DPCCA	69.77	0.0181

The reported interpretation is that D$\mathbf{z}^1,\mathbf{z}^2$9PCCA substantially improves ELBO over linear DPCCA, and that KL annealing and normalizing flows further improve ELBO. The RMSE differences are small, which the paper interprets as evidence that the main gain lies in probabilistic fit and latent dynamic modeling rather than only pointwise reconstruction. Qualitative reconstruction plots with uncertainty bands show higher variance near the beginning of sequences and nonlinear temporal trends in stock prices, with D$^2$0PCCA better capturing these than DPCCA.

The abstract further states that D$^2$1PCCA naturally extends to multiple observed variables, making it a versatile tool for encoding prior knowledge about sequential datasets and providing a probabilistic understanding of the system’s dynamics. Taken together, these results position D$^2$2PCCA as a probabilistic, deep nonlinear dynamic generalization of PFCCA-style ideas: it preserves the two-view shared/private decomposition of CCA, adds temporal dynamics, performs past-future inference through dependence on future observations, and supports uncertainty quantification through a generative latent-state framework.