Time-Causal Wavelet Analysis

• Time-causal wavelet analysis is a set of wavelet techniques that use strictly causal, one-sided kernels (like truncated exponentials and causalized Morlet functions) to analyze temporal signals using only past data.
• It combines multiscale localization with recursive implementation to ensure temporal scale covariance, balancing frequency resolution with controlled delay.
• The method has practical applications in early warning systems, such as detecting power buildup and phase shifts in environmental time series like ozone concentration.

Searching arXiv for recent and foundational papers on time-causal wavelet analysis. Time-causal wavelet analysis is a class of wavelet methods for temporal signals in which every stage of the analysis is constrained to be strictly causal, so that coefficients at time $u$ depend only on data from times $t \le u$. In this setting, the usual symmetric wavelet constructions are replaced by one-sided kernels, typically based on truncated exponentials in cascade or on causalized generalized Morlet constructions, in order to support real-time processing without look-ahead. The resulting framework combines multiscale localization, recursive implementation, and physically realistic temporal support, and has been developed both as a general theory linked to temporal scale-space and as a practical methodology for early warning in non-stationary environmental time series (Lindeberg, 7 Oct 2025, Martínez-Cadena et al., 2024).

1. Definition and scope

Time-causal wavelet analysis addresses a basic incompatibility between standard wavelet analysis and real-time temporal inference. Conventional wavelet transforms commonly rely on symmetric kernels, so future and past samples contribute jointly to the coefficient at a given time. For streaming signals, online monitoring, and physically causal temporal processes, that construction is not admissible. In the time-causal formulation, the wavelet kernel has support only on the causal half-axis, and the transform is therefore defined using only the observed past (Lindeberg, 7 Oct 2025).

A central distinction is between non-causal and causal localization functions. In the generalized Morlet construction used for ozone analysis, the non-causal wavelet begins from the Ansatz

$\psi(\sigma)=M(\sigma)e^{i\sigma},$

with dimensionless coordinate $\sigma=(t-u)/s$, where $M(\sigma)$ is a real-valued modulation function. Martínez-Cadena et al. choose the one-parameter Mittag–Leffler function as a generalization of the Gaussian,

$$M(\sigma)=E_\alpha(-|\sigma|^\alpha) =\sum_{k=0}^\infty \frac{(-|\sigma|^\alpha)^k}{\Gamma(1+k\alpha)}.$$

Strict causality is then enforced by truncating the modulation to the past,

$$M_{\text{causal}}(\sigma)= \begin{cases} E_\alpha(-\sigma^\alpha), & \sigma\ge 0,\ 0, & \sigma<0. \end{cases}$$

This yields a causal wavelet with support only for $t\le u$, with a factor $1/\sqrt{s}$ to normalize energy across scales (Martínez-Cadena et al., 2024).

A broader theoretical formulation replaces causalized Morlet envelopes by temporal derivatives of a time-causal limit kernel from temporal scale-space theory. In that construction, mother wavelets are generated from

$\psi_n(t)=(-1)^n\,\partial_t^n \Psi(t;\tau_0,c),$

where $t \le u$0 is the limit of an infinite cascade of truncated exponentials selected to satisfy temporal scale covariance under $t \le u$1-adic scaling (Lindeberg, 7 Oct 2025). This suggests that “time-causal wavelet analysis” is not a single kernel family but a methodological class defined by one-sided support, recursive realizability, and scale-space compatibility.

2. Kernel constructions and admissibility

The generalized Morlet formulation and the temporal scale-space formulation provide two complementary constructions. In the former, one starts from a complex sinusoid modulated by a one-sided localization function. In the latter, one derives admissible wavelets from causal smoothing kernels and their temporal derivatives. Both formulations are explicitly designed to avoid future influence (Martínez-Cadena et al., 2024, Lindeberg, 7 Oct 2025).

In the causal generalized Morlet case, the wavelet shifted to time $t \le u$2 and dilated by scale $t \le u$3 is

$t \le u$4

For $t \le u$5, or through the stretched-exponential approximation $t \le u$6, one recovers the classical Morlet Gaussian window. Martínez-Cadena et al. use $t \le u$7, that is, a causal Morlet choice (Martínez-Cadena et al., 2024).

In the scale-space formulation, the admissible smoothing primitives are more tightly constrained. The theory requires that smoothing from finer to coarser scales never increases the number of zero-crossings or local extrema. On a causal time axis, the only primitive variation-diminishing kernels are the causal exponentials

$t \le u$8

A multi-scale representation is therefore built by cascading first-order integrators, and any other choice would either require access to future values or would violate variation-diminishing (Lindeberg, 7 Oct 2025).

The corresponding limit kernel has Laplace transform

$t \le u$9

with time constants chosen as

$\psi(\sigma)=M(\sigma)e^{i\sigma},$0

so that $\psi(\sigma)=M(\sigma)e^{i\sigma},$1 is exactly scale-covariant under $\psi(\sigma)=M(\sigma)e^{i\sigma},$2 and $\psi(\sigma)=M(\sigma)e^{i\sigma},$3 (Lindeberg, 7 Oct 2025).

Admissibility follows in the derivative-based construction from zero mean and fast temporal decay. Specifically,

$\psi(\sigma)=M(\sigma)e^{i\sigma},$4

and the exponential tails imply integrability and a finite admissibility constant. The Fourier transform is

$\psi(\sigma)=M(\sigma)e^{i\sigma},$5

A plausible implication is that the derivative-based family plays the role that derivative-of-Gaussian and Morlet families play in non-causal wavelet analysis, but under a strictly one-sided temporal support constraint (Lindeberg, 7 Oct 2025).

3. Transform definitions, power, and phase

Given a real-valued signal $\psi(\sigma)=M(\sigma)e^{i\sigma},$6, the time-causal wavelet coefficients in the Martínez-Cadena et al. formulation are

$\psi(\sigma)=M(\sigma)e^{i\sigma},$7

where $\psi(\sigma)=M(\sigma)e^{i\sigma},$8. Equivalently, with

$\psi(\sigma)=M(\sigma)e^{i\sigma},$9

one writes

$\sigma=(t-u)/s$0

This is a one-sided convolution-like operation over the causal past (Martínez-Cadena et al., 2024).

Local power density is defined as

$\sigma=(t-u)/s$1

and the image of $\sigma=(t-u)/s$2 over time $\sigma=(t-u)/s$3 and scales $\sigma=(t-u)/s$4 is the scalogram. A global scale spectrum may be obtained by averaging in time,

$\sigma=(t-u)/s$5

In a strictly causal setting, a high-power event at a small scale $\sigma=(t-u)/s$6 can leak into larger scales $\sigma=(t-u)/s$7 as the fading-memory kernel integrates past contributions. In the time-scale plane this appears as vertical ridges in the scalogram and is interpreted as local accumulation (Martínez-Cadena et al., 2024).

Phase carries an additional diagnostic role. Writing

$\sigma=(t-u)/s$8

one defines

$\sigma=(t-u)/s$9

In the ozone study, $M(\sigma)$0 corresponds to dominance of the even (cosine) component and is interpreted as anti-persistence, whereas $M(\sigma)$1 corresponds to dominance of the odd (sine) component and is interpreted as persistence. Ramps in power together with a phase moving toward zero can signal the onset of anti-persistent build-up (Martínez-Cadena et al., 2024).

The derivative-based scale-space formulation expresses the wavelet transform over discrete scales $M(\sigma)$2 as

$M(\sigma)$3

This formulation emphasizes the link between temporal scale covariance and discrete scale sampling. Finer sampling of scales corresponds to smaller $M(\sigma)$4, for example $M(\sigma)$5, which yields a denser wavelet bank but larger group delay (Lindeberg, 7 Oct 2025).

4. Time-recursive implementation

A defining feature of time-causal wavelet analysis is that the kernels admit recursive implementations. In practice, the causal wavelet may be realized as a one-sided fading-memory filter. A common implementation is a cascade of $M(\sigma)$6 first-order recursive exponential filters, each causal, whose combined impulse response is a Gamma kernel (Martínez-Cadena et al., 2024).

For a single first-order causal filter with time constant $M(\sigma)$7, the impulse response is

$M(\sigma)$8

Cascading $M(\sigma)$9 such filters with the same $$M(\sigma)=E_\alpha(-|\sigma|^\alpha) =\sum_{k=0}^\infty \frac{(-|\sigma|^\alpha)^k}{\Gamma(1+k\alpha)}.$$0 yields the $$M(\sigma)=E_\alpha(-|\sigma|^\alpha) =\sum_{k=0}^\infty \frac{(-|\sigma|^\alpha)^k}{\Gamma(1+k\alpha)}.$$1th-order Gamma kernel

$$M(\sigma)=E_\alpha(-|\sigma|^\alpha) =\sum_{k=0}^\infty \frac{(-|\sigma|^\alpha)^k}{\Gamma(1+k\alpha)}.$$2

with Fourier transform

$$M(\sigma)=E_\alpha(-|\sigma|^\alpha) =\sum_{k=0}^\infty \frac{(-|\sigma|^\alpha)^k}{\Gamma(1+k\alpha)}.$$3

To approximate the generalized Morlet with $$M(\sigma)=E_\alpha(-|\sigma|^\alpha) =\sum_{k=0}^\infty \frac{(-|\sigma|^\alpha)^k}{\Gamma(1+k\alpha)}.$$4, one chooses $$M(\sigma)=E_\alpha(-|\sigma|^\alpha) =\sum_{k=0}^\infty \frac{(-|\sigma|^\alpha)^k}{\Gamma(1+k\alpha)}.$$5 large and $$M(\sigma)=E_\alpha(-|\sigma|^\alpha) =\sum_{k=0}^\infty \frac{(-|\sigma|^\alpha)^k}{\Gamma(1+k\alpha)}.$$6 so that $$M(\sigma)=E_\alpha(-|\sigma|^\alpha) =\sum_{k=0}^\infty \frac{(-|\sigma|^\alpha)^k}{\Gamma(1+k\alpha)}.$$7 for $$M(\sigma)=E_\alpha(-|\sigma|^\alpha) =\sum_{k=0}^\infty \frac{(-|\sigma|^\alpha)^k}{\Gamma(1+k\alpha)}.$$8, then modulates by the complex sinusoid $$M(\sigma)=E_\alpha(-|\sigma|^\alpha) =\sum_{k=0}^\infty \frac{(-|\sigma|^\alpha)^k}{\Gamma(1+k\alpha)}.$$9 and normalizes by $$M_{\text{causal}}(\sigma)= \begin{cases} E_\alpha(-\sigma^\alpha), & \sigma\ge 0,\ 0, & \sigma<0. \end{cases}$$0 (Martínez-Cadena et al., 2024).

The more general time-recursive implementation truncates the infinite cascade at $$M_{\text{causal}}(\sigma)= \begin{cases} E_\alpha(-\sigma^\alpha), & \sigma\ge 0,\ 0, & \sigma<0. \end{cases}$$1 stages. If $$M_{\text{causal}}(\sigma)= \begin{cases} E_\alpha(-\sigma^\alpha), & \sigma\ge 0,\ 0, & \sigma<0. \end{cases}$$2 is chosen so that

$$M_{\text{causal}}(\sigma)= \begin{cases} E_\alpha(-\sigma^\alpha), & \sigma\ge 0,\ 0, & \sigma<0. \end{cases}$$3

with the first $$M_{\text{causal}}(\sigma)= \begin{cases} E_\alpha(-\sigma^\alpha), & \sigma\ge 0,\ 0, & \sigma<0. \end{cases}$$4 absorbing the variance of all omitted deeper layers, then each scale channel is updated recursively as

$$M_{\text{causal}}(\sigma)= \begin{cases} E_\alpha(-\sigma^\alpha), & \sigma\ge 0,\ 0, & \sigma<0. \end{cases}$$5

After $$M_{\text{causal}}(\sigma)= \begin{cases} E_\alpha(-\sigma^\alpha), & \sigma\ge 0,\ 0, & \sigma<0. \end{cases}$$6 cascaded updates, one obtains a discrete approximation of $$M_{\text{causal}}(\sigma)= \begin{cases} E_\alpha(-\sigma^\alpha), & \sigma\ge 0,\ 0, & \sigma<0. \end{cases}$$7. Temporal derivatives are then approximated by finite differences such as

$$M_{\text{causal}}(\sigma)= \begin{cases} E_\alpha(-\sigma^\alpha), & \sigma\ge 0,\ 0, & \sigma<0. \end{cases}$$8

followed by multiplication by $$M_{\text{causal}}(\sigma)= \begin{cases} E_\alpha(-\sigma^\alpha), & \sigma\ge 0,\ 0, & \sigma<0. \end{cases}$$9 for scale-normalization (Lindeberg, 7 Oct 2025).

Because the filter bank is strictly causal and recursive, no buffer beyond the $t\le u$0 channel states is needed. The applications explicitly listed include hardware or fixed-latency DSP, physiological signals such as EEG and ECG, industrial-control loops, robotics, and online monitoring (Lindeberg, 7 Oct 2025). This suggests that the computational architecture is as central to the framework as the wavelet definition itself.

5. Scale covariance, temporal duration, and scale selection

The scale-space formulation emphasizes exact temporal scale covariance for a specific distribution of time constants. Under $t\le u$1 and $t\le u$2 with $t\le u$3,

$t\le u$4

and the $t\le u$5th-order derivative obeys

$t\le u$6

These identities formalize the self-similarity of the causal representation across temporal scales (Lindeberg, 7 Oct 2025).

Temporal duration grows in a controlled way with scale. The variance of $t\le u$7 is exactly $t\le u$8, so its stretch grows like $t\le u$9. The mean delay is given by

$1/\sqrt{s}$0

and the peak delay $1/\sqrt{s}$1 can be approximated in closed form via Koenderink’s formula; both scale like $1/\sqrt{s}$2. The paper further states that smaller $1/\sqrt{s}$3 gives denser scale sampling but larger group delay, making delay-scale trade-offs explicit rather than incidental (Lindeberg, 7 Oct 2025).

Scale selection can then be performed from extrema over scales. If a signal contains a blob of width $1/\sqrt{s}$4 modeled by a Gaussian $1/\sqrt{s}$5, then its strongest response under the scale-normalized second derivative with normalization power $1/\sqrt{s}$6 occurs at $1/\sqrt{s}$7. Similarly, an edge of rise-time $1/\sqrt{s}$8 is selected by the first derivative with $1/\sqrt{s}$9 at $\psi_n(t)=(-1)^n\,\partial_t^n \Psi(t;\tau_0,c),$0. In practice, one scans over discrete $\psi_n(t)=(-1)^n\,\partial_t^n \Psi(t;\tau_0,c),$1 and finds

$\psi_n(t)=(-1)^n\,\partial_t^n \Psi(t;\tau_0,c),$2

This provides a wavelet-style mechanism for estimating the duration of locally dominant temporal structures in a strictly causal representation (Lindeberg, 7 Oct 2025).

In the ozone application, scale selection is operational rather than axiomatic. Scales $\psi_n(t)=(-1)^n\,\partial_t^n \Psi(t;\tau_0,c),$3 range from 2 days up to approximately 900 days in geometric progression, with approximately 40 scales and time step $\psi_n(t)=(-1)^n\,\partial_t^n \Psi(t;\tau_0,c),$4 day. The example discussed in detail uses $\psi_n(t)=(-1)^n\,\partial_t^n \Psi(t;\tau_0,c),$5 days for warning analysis, illustrating how a specific temporal horizon can be read off from the multiscale representation (Martínez-Cadena et al., 2024).

6. Empirical application to ozone contingencies and methodological limits

Martínez-Cadena et al. apply the causal generalized Morlet methodology to daily-averaged ozone concentration in the Mexico City Metropolitan Area from 2010–2023, comprising 4,899 points, with gaps of less than 0.5% of records imputed per Diosdado et al. (2013). Contingency days are flagged when $\psi_n(t)=(-1)^n\,\partial_t^n \Psi(t;\tau_0,c),$6 IMECAS. The parameter choices are modulation exponent $\psi_n(t)=(-1)^n\,\partial_t^n \Psi(t;\tau_0,c),$7, a filter cascade with $\psi_n(t)=(-1)^n\,\partial_t^n \Psi(t;\tau_0,c),$8–$\psi_n(t)=(-1)^n\,\partial_t^n \Psi(t;\tau_0,c),$9 to approximate Gaussian behavior, $t \le u$00, scales from 2 days to about 900 days in geometric progression, and $t \le u$01 day (Martínez-Cadena et al., 2024).

The causal scalogram of $t \le u$02 shows strong vertical bands of power buildup 10–50 days before many declared contingencies. In the same lead-time window, the phase $t \le u$03 dips toward zero, which is interpreted as anti-persistence. The authors conclude that dangerous ozone concentration levels are a consequence of accumulation and incomplete dissipation effects acting over a wide range of time scales, and that contingencies occurred when the wavelet coefficient power is increasing, linked to an anti-persistence behavior (Martínez-Cadena et al., 2024).

A specific operational example is reported at scale $t \le u$04 days: $t \le u$05 rising past a normalized threshold, for example 0.5 of max power, gave 14 warning flags over 2010–2023, many preceding actual emergency days. The proposed interpretation is that joint monitoring of power build-up and phase orientation can serve as a pre-contingency warning signal (Martínez-Cadena et al., 2024).

The framework’s stated advantages are real-time applicability, because no future data are needed; multiscale insight, because one transform monitors short- and long-term buildup simultaneously; and phase-power joint analysis, because it distinguishes accumulation from persistence type. The stated limitations are that the choice of $t \le u$06, $t \le u$07, and $t \le u$08 trades temporal localization against frequency resolution and requires tuning; boundary effects arise at the start of the record because purely causal kernels have no pre-data; very long-scale responses may be dominated by early-record transients unless warm-up is discarded; and high-$t \le u$09 causal cascades can have numerical cost when extremely fine scale sampling is needed (Martínez-Cadena et al., 2024).

A common misconception is that causality merely requires truncating a standard wavelet. The general theory indicates a stronger claim: on a causal time axis, truncated exponentials in cascade constitute the only permissable class of kernels for variation-diminishing temporal scale-space, with temporal derivatives as the natural complement to satisfy wavelet admissibility conditions (Lindeberg, 7 Oct 2025). This does not invalidate causalized Morlet constructions, but it places them within a broader design space in which scale-space constraints, delay, and admissibility must be considered jointly.