Time-Causal Limit Kernel

• Time-Causal Limit Kernel is a temporal smoothing kernel built as the limit of cascaded truncated exponential filters, offering discrete scale invariance for real-time analysis.
• It enforces strict time-causality and a variation-diminishing property, ensuring that no new local extrema or zero-crossings are introduced with increasing temporal scale.
• Its recursive implementation via a cascade structure enables efficient, real-time processing in applications such as receptive-field modeling, feature detection, and time-causal wavelet analysis.

The time-causal limit kernel is a temporal smoothing kernel for real-time multi-scale analysis that is defined as the limit of a cascade of truncated exponential filters with logarithmically or geometrically distributed time constants. In the cited scale-space formulations, it is the canonical kernel obtained when strict time-causality, time-recursiveness, and non-creation of new local extrema or zero-crossings with increasing temporal scale are imposed over a one-dimensional temporal domain. Its defining property is exact self-similarity under a discrete set of temporal rescalings, which yields true temporal scale invariance for scaling factors matched to a geometric scale ladder, while preserving a variation-diminishing cascade structure and enabling recursive implementations with finite state (Lindeberg, 2015, Lindeberg, 2017, Lindeberg, 2022).

1. Axiomatic basis and conceptual status

The time-causal limit kernel arises in temporal scale-space theory from a different axiomatic setting than the non-causal Gaussian case. Over time, the primary requirement is not a continuous semi-group over scales, but non-creation of new structures—specifically, non-creation of new local extrema or zero-crossings—with increasing temporal scale. Time causality further enforces that the temporal kernel satisfies $h(t;\tau)=0$ for $t<0$ (Lindeberg, 2015).

This framework is anchored in the theory of variation-diminishing linear transformations associated with Schoenberg and Karlin. In the temporal case, time-causal kernels that do not create new extrema must be generated by truncated exponential filters. The cited works state that, under temporal scale-space axioms and strict time-causality, the only admissible smoothing kernels are truncated exponentials, and multi-scale smoothing must be built by cascading them (Lindeberg, 2015, Lindeberg, 7 Oct 2025).

A central consequence is that the temporal scale parameter is semi-discrete rather than continuous. The weaker cascade or Markov property is retained:

$$h(\cdot;\; \tau_2) = (\Delta h)(\cdot;\; \tau_1 \mapsto \tau_2) * h(\cdot;\; \tau_1), \quad \tau_2 > \tau_1,$$

where the transformation kernels are themselves temporal scale-space kernels (Lindeberg, 2015). This distinguishes the construction from the Gaussian temporal kernel, which is non-causal, symmetric, continuous in scale, and obeys a strict semi-group over $\tau$ (Lindeberg, 2017).

A common misconception is that strict time-causality can be combined with the same continuous semi-group structure that characterizes Gaussian scale space. In the cited theory, this is explicitly not the case: the temporal scale parameter must be discrete, and exact temporal scale invariance is available only for scale mappings that correspond to integer powers of the base $c$ (Lindeberg, 2015, Lindeberg, 2017).

2. Construction from truncated exponentials

The primitive building block is the truncated exponential, equivalently the impulse response of a first-order integrator:

$$h_{exp}(t;\; \mu_k) = \begin{cases} \tfrac{1}{\mu_k}\, e^{-t/\mu_k}, & t \ge 0,\ 0, & t<0, \end{cases}$$

with Laplace transform

$H_{exp}(q;\; \mu_k) = \frac{1}{1 + \mu_k q}.$

Coupling $K$ such filters in cascade gives

$$h_{composed}(\cdot;\; \mu) = *_{k=1}^{K} h_{exp}(\cdot;\; \mu_k), \quad H_{composed}(q;\; \mu) = \prod_{k=1}^{K} \frac{1}{1 + \mu_k q},$$

and the mean and variance are additive under convolution:

$$m_K = \sum_{k=1}^{K} \mu_k, \qquad \tau_K = \sum_{k=1}^{K} \mu_k^2.$$

In continuous time, repeated convolution with truncated exponentials corresponds to coupling first-order integrators in cascade:

$t<0$0

(Lindeberg, 2015).

Three temporal scale parameterizations are described for the finite cascade. For a uniform distribution of intermediate scales,

$t<0$1

which yields equal time constants,

$t<0$2

For a logarithmic or geometric distribution with free minimum scale,

$t<0$3

with

$t<0$4

A third variant fixes $t<0$5 by setting

$t<0$6

These distributions are used to tune temporal dynamics and to approach the scale-invariant limit (Lindeberg, 2015).

With logarithmically distributed intermediate scales, the time-causal limit kernel is obtained by letting $t<0$7. Its Fourier transform is

$t<0$8

Equivalent Laplace-domain and cascade representations are also given:

$t<0$9

with total variance

$$h(\cdot;\; \tau_2) = (\Delta h)(\cdot;\; \tau_1 \mapsto \tau_2) * h(\cdot;\; \tau_1), \quad \tau_2 > \tau_1,$$0

(Lindeberg, 2015, Lindeberg, 7 Oct 2025, Lindeberg, 2022).

The cited works also provide a series representation by partial fractions:

$$h(\cdot;\; \tau_2) = (\Delta h)(\cdot;\; \tau_1 \mapsto \tau_2) * h(\cdot;\; \tau_1), \quad \tau_2 > \tau_1,$$1

and for derivatives

$$h(\cdot;\; \tau_2) = (\Delta h)(\cdot;\; \tau_1 \mapsto \tau_2) * h(\cdot;\; \tau_1), \quad \tau_2 > \tau_1,$$2

(Lindeberg, 7 Oct 2025).

3. Self-similarity, scale covariance, and temporal dynamics

The defining theoretical feature of the time-causal limit kernel is exact self-similarity across a discrete family of temporal scales. The recursive relation between adjacent scales is

$$h(\cdot;\; \tau_2) = (\Delta h)(\cdot;\; \tau_1 \mapsto \tau_2) * h(\cdot;\; \tau_1), \quad \tau_2 > \tau_1,$$3

or in the Fourier domain,

$$h(\cdot;\; \tau_2) = (\Delta h)(\cdot;\; \tau_1 \mapsto \tau_2) * h(\cdot;\; \tau_1), \quad \tau_2 > \tau_1,$$4

The limit kernel further obeys exact scale invariance under temporal rescaling:

$$h(\cdot;\; \tau_2) = (\Delta h)(\cdot;\; \tau_1 \mapsto \tau_2) * h(\cdot;\; \tau_1), \quad \tau_2 > \tau_1,$$5

so that rescaling the signal by $$h(\cdot;\; \tau_2) = (\Delta h)(\cdot;\; \tau_1 \mapsto \tau_2) * h(\cdot;\; \tau_1), \quad \tau_2 > \tau_1,$$6 maps the scale-space at $$h(\cdot;\; \tau_2) = (\Delta h)(\cdot;\; \tau_1 \mapsto \tau_2) * h(\cdot;\; \tau_1), \quad \tau_2 > \tau_1,$$7 to that at $$h(\cdot;\; \tau_2) = (\Delta h)(\cdot;\; \tau_1 \mapsto \tau_2) * h(\cdot;\; \tau_1), \quad \tau_2 > \tau_1,$$8:

$$h(\cdot;\; \tau_2) = (\Delta h)(\cdot;\; \tau_1 \mapsto \tau_2) * h(\cdot;\; \tau_1), \quad \tau_2 > \tau_1,$$9

Applied recursively, the representation is closed under all temporal scalings $\tau$0, $\tau$1 (Lindeberg, 2015).

This discrete covariance property is the basis for the claim of true temporal scale invariance in a time-causal setting. The cited works are explicit that exact covariance holds for scaling factors aligned with the geometric scale ladder, while for other rescalings deviations are small when $\tau$2 is close to $\tau$3, for example $\tau$4 (Lindeberg, 2017, Lindeberg, 7 Oct 2025).

The logarithmic distribution of intermediate scales is introduced partly because it leads to faster temporal response than a uniform distribution. For equal-$\tau$5 filters, the mean delay is

$\tau$6

whereas for logarithmically spaced $\tau$7

$\tau$8

with limit

$\tau$9

For the uniform case, the position of the local maximum is

$c$0

and numerical comparisons show substantial reductions in $c$1 under logarithmic spacing, especially as $c$2 grows (Lindeberg, 2015).

For the limit kernel, the moments and cumulants are given in closed form. The mean and variance satisfy

$c$3

Higher-order terms include

$c$4

and the cited analyses state that skewness and kurtosis are strictly positive and increase with $c$5 (Lindeberg, 2015). By contrast, for uniform spacing of intermediate scales,

$c$6

which tend to zero as $c$7 grows, so the uniform cascade converges to a symmetric Gaussian-like limit, whereas the logarithmic cascade converges to a genuinely skewed limit distribution (Lindeberg, 2015).

In the frequency domain, the magnitude, phase, and group delay are

$c$8

$c$9

$$h_{exp}(t;\; \mu_k) = \begin{cases} \tfrac{1}{\mu_k}\, e^{-t/\mu_k}, & t \ge 0,\ 0, & t<0, \end{cases}$$0

As $$h_{exp}(t;\; \mu_k) = \begin{cases} \tfrac{1}{\mu_k}\, e^{-t/\mu_k}, & t \ge 0,\ 0, & t<0, \end{cases}$$1, the delay grows with $$h_{exp}(t;\; \mu_k) = \begin{cases} \tfrac{1}{\mu_k}\, e^{-t/\mu_k}, & t \ge 0,\ 0, & t<0, \end{cases}$$2; as $$h_{exp}(t;\; \mu_k) = \begin{cases} \tfrac{1}{\mu_k}\, e^{-t/\mu_k}, & t \ge 0,\ 0, & t<0, \end{cases}$$3 increases, the delay diminishes (Lindeberg, 2017).

4. Scale-normalized derivatives, feature detection, and wavelets

The time-causal limit kernel is not only a smoothing kernel but also the basis for scale-normalized derivative constructions. Over time, two normalizations are described. The variance-based normalization is

$$h_{exp}(t;\; \mu_k) = \begin{cases} \tfrac{1}{\mu_k}\, e^{-t/\mu_k}, & t \ge 0,\ 0, & t<0, \end{cases}$$4

while the $$h_{exp}(t;\; \mu_k) = \begin{cases} \tfrac{1}{\mu_k}\, e^{-t/\mu_k}, & t \ge 0,\ 0, & t<0, \end{cases}$$5-normalization chooses $$h_{exp}(t;\; \mu_k) = \begin{cases} \tfrac{1}{\mu_k}\, e^{-t/\mu_k}, & t \ge 0,\ 0, & t<0, \end{cases}$$6 such that

$$h_{exp}(t;\; \mu_k) = \begin{cases} \tfrac{1}{\mu_k}\, e^{-t/\mu_k}, & t \ge 0,\ 0, & t<0, \end{cases}$$7

For spatio-temporal derivatives of spatial order $$h_{exp}(t;\; \mu_k) = \begin{cases} \tfrac{1}{\mu_k}\, e^{-t/\mu_k}, & t \ge 0,\ 0, & t<0, \end{cases}$$8 and temporal order $$h_{exp}(t;\; \mu_k) = \begin{cases} \tfrac{1}{\mu_k}\, e^{-t/\mu_k}, & t \ge 0,\ 0, & t<0, \end{cases}$$9,

$H_{exp}(q;\; \mu_k) = \frac{1}{1 + \mu_k q}.$0

(Lindeberg, 2015).

For the limit kernel, normalized temporal derivatives are exactly scale invariant or covariant under the admissible discrete scaling group. The transformation law is

$H_{exp}(q;\; \mu_k) = \frac{1}{1 + \mu_k q}.$1

with full invariance for $H_{exp}(q;\; \mu_k) = \frac{1}{1 + \mu_k q}.$2:

$H_{exp}(q;\; \mu_k) = \frac{1}{1 + \mu_k q}.$3

The cited papers treat $H_{exp}(q;\; \mu_k) = \frac{1}{1 + \mu_k q}.$4 as the normalization that enforces scale invariance across mapped scales when the limit kernel is used (Lindeberg, 2015, Lindeberg, 2017).

These derivative constructions support temporal scale selection by local extrema over temporal scales of scale-normalized derivative responses. For the time-causal limit kernel, the papers state that maxima over $H_{exp}(q;\; \mu_k) = \frac{1}{1 + \mu_k q}.$5 of such responses preserve characteristic temporal durations under rescaling, and that for a sine wave the selected scale is proportional to wavelength:

$H_{exp}(q;\; \mu_k) = \frac{1}{1 + \mu_k q}.$6

For $H_{exp}(q;\; \mu_k) = \frac{1}{1 + \mu_k q}.$7, the maximum scale-normalized magnitude is independent of $H_{exp}(q;\; \mu_k) = \frac{1}{1 + \mu_k q}.$8, which is the scale-invariant magnitude property described in the paper (Lindeberg, 2017).

The same kernel underlies later wavelet constructions. In the time-causal wavelet formulation, mother wavelets are defined as temporal derivatives of the limit kernel:

$H_{exp}(q;\; \mu_k) = \frac{1}{1 + \mu_k q}.$9

They satisfy zero integral,

$K$0

and the admissibility condition

$K$1

because the series representation yields exponential decay in $K$2 and $K$3 near $K$4 (Lindeberg, 7 Oct 2025).

A plausible implication is that the time-causal limit kernel functions as a common substrate for temporal scale-space, derivative-based feature detection, and strictly time-causal wavelet analysis, because all three constructions reuse the same variation-diminishing and scale-covariant cascade structure (Lindeberg, 2015, Lindeberg, 7 Oct 2025).

5. Discrete-time realization and recursive computation

A major motivation for the time-causal limit kernel is operational suitability for sampled data and real-time systems. For sampled video, the temporal scale is related to frame rate by

$K$5

Each stage in the discrete-time realization uses a first-order recursive filter:

$K$6

with generating function

$K$7

and

$K$8

where $K$9 (Lindeberg, 2015).

The composed discrete filter is

$$h_{composed}(\cdot;\; \mu) = *_{k=1}^{K} h_{exp}(\cdot;\; \mu_k), \quad H_{composed}(q;\; \mu) = \prod_{k=1}^{K} \frac{1}{1 + \mu_k q},$$0

The cited implementation requires only $$h_{composed}(\cdot;\; \mu) = *_{k=1}^{K} h_{exp}(\cdot;\; \mu_k), \quad H_{composed}(q;\; \mu) = \prod_{k=1}^{K} \frac{1}{1 + \mu_k q},$$1 scale-buffered states and no explicit storage of history. It is therefore time-recursive in the specific sense that the memory of the past is represented by the current scale-space states themselves rather than by a long temporal buffer (Lindeberg, 2015, Lindeberg, 2022).

Equivalent update forms are also given in later treatments. One recurrent formulation is

$$h_{composed}(\cdot;\; \mu) = *_{k=1}^{K} h_{exp}(\cdot;\; \mu_k), \quad H_{composed}(q;\; \mu) = \prod_{k=1}^{K} \frac{1}{1 + \mu_k q},$$2

while another writes

$$h_{composed}(\cdot;\; \mu) = *_{k=1}^{K} h_{exp}(\cdot;\; \mu_k), \quad H_{composed}(q;\; \mu) = \prod_{k=1}^{K} \frac{1}{1 + \mu_k q},$$3

The papers describe these as stable first-order IIR low-pass stages in cascade and emphasize that the algorithm has $$h_{composed}(\cdot;\; \mu) = *_{k=1}^{K} h_{exp}(\cdot;\; \mu_k), \quad H_{composed}(q;\; \mu) = \prod_{k=1}^{K} \frac{1}{1 + \mu_k q},$$4 operations and $$h_{composed}(\cdot;\; \mu) = *_{k=1}^{K} h_{exp}(\cdot;\; \mu_k), \quad H_{composed}(q;\; \mu) = \prod_{k=1}^{K} \frac{1}{1 + \mu_k q},$$5 states per sample (Lindeberg, 7 Oct 2025, Lindeberg, 2023).

For finite approximations to the limit kernel, a recommended variance-preserving distribution is

$$h_{composed}(\cdot;\; \mu) = *_{k=1}^{K} h_{exp}(\cdot;\; \mu_k), \quad H_{composed}(q;\; \mu) = \prod_{k=1}^{K} \frac{1}{1 + \mu_k q},$$6

ensuring

$$h_{composed}(\cdot;\; \mu) = *_{k=1}^{K} h_{exp}(\cdot;\; \mu_k), \quad H_{composed}(q;\; \mu) = \prod_{k=1}^{K} \frac{1}{1 + \mu_k q},$$7

The convergence to the limit is described as rapid, and the practical guidance repeatedly states that $$h_{composed}(\cdot;\; \mu) = *_{k=1}^{K} h_{exp}(\cdot;\; \mu_k), \quad H_{composed}(q;\; \mu) = \prod_{k=1}^{K} \frac{1}{1 + \mu_k q},$$8–$$h_{composed}(\cdot;\; \mu) = *_{k=1}^{K} h_{exp}(\cdot;\; \mu_k), \quad H_{composed}(q;\; \mu) = \prod_{k=1}^{K} \frac{1}{1 + \mu_k q},$$9 typically suffices (Lindeberg, 7 Oct 2025, Lindeberg, 2023).

The same recursive architecture extends naturally to time-causal time-frequency analysis. In the time-causal analogue of the Gabor transform, the Gaussian window is replaced by $$m_K = \sum_{k=1}^{K} \mu_k, \qquad \tau_K = \sum_{k=1}^{K} \mu_k^2.$$0, yielding

$$m_K = \sum_{k=1}^{K} \mu_k, \qquad \tau_K = \sum_{k=1}^{K} \mu_k^2.$$1

and

$$m_K = \sum_{k=1}^{K} \mu_k, \qquad \tau_K = \sum_{k=1}^{K} \mu_k^2.$$2

This forward transform remains strictly causal and inherits the same cascade property over temporal scales (Lindeberg, 2023).

6. Receptive-field modeling, applications, and limitations

The original 2015 formulation embeds the time-causal limit kernel in a separable spatio-temporal model,

$$m_K = \sum_{k=1}^{K} \mu_k, \qquad \tau_K = \sum_{k=1}^{K} \mu_k^2.$$3

with a spatial Gaussian or affine Gaussian factor and a time-causal temporal kernel (Lindeberg, 2015). This leads to receptive-field models for both LGN and V1. For LGN,

$$m_K = \sum_{k=1}^{K} \mu_k, \qquad \tau_K = \sum_{k=1}^{K} \mu_k^2.$$4

where $$m_K = \sum_{k=1}^{K} \mu_k, \qquad \tau_K = \sum_{k=1}^{K} \mu_k^2.$$5 for non-lagged and $$m_K = \sum_{k=1}^{K} \mu_k, \qquad \tau_K = \sum_{k=1}^{K} \mu_k^2.$$6 for lagged cells. For V1 simple cells,

$$m_K = \sum_{k=1}^{K} \mu_k, \qquad \tau_K = \sum_{k=1}^{K} \mu_k^2.$$7

capturing orientation, velocity tuning, and subfield structure (Lindeberg, 2015).

The same framework supports feature detectors based on partial derivatives, directional derivatives, spatio-temporal invariants such as the determinant of the $$m_K = \sum_{k=1}^{K} \mu_k, \qquad \tau_K = \sum_{k=1}^{K} \mu_k^2.$$8D Hessian and rescaled $$m_K = \sum_{k=1}^{K} \mu_k, \qquad \tau_K = \sum_{k=1}^{K} \mu_k^2.$$9D Gaussian curvature, and quasi-quadrature energy measures that quantify joint odd/even energy across space-time and support phase-insensitive detection of motion, flicker, and transients (Lindeberg, 2015).

In later work, the time-causal limit kernel is also used for temporal scale selection, time-causal wavelets, and a time-causal analogue of the Gabor transform. These developments all preserve the central claims: strict causality, recursive computation, a cascade property over discrete scales, and exact temporal scale covariance for the discrete scaling group $t<0$00 (Lindeberg, 2017, Lindeberg, 2023, Lindeberg, 7 Oct 2025).

Several limitations and trade-offs are explicit in the cited literature. The temporal scale parameter must be discrete, and full temporal scale invariance is available only at scale mappings that correspond to integer powers of the base $t<0$01 (Lindeberg, 2015). The base $t<0$02 controls a trade-off between sampling density and speed or skewness: smaller $t<0$03 yields denser scales, while larger $t<0$04 produces faster responses and more skewed kernels (Lindeberg, 2015). Practical guidance is to choose $t<0$05 or, in later formulations, $t<0$06 when scale accuracy is primary and $t<0$07 when responsiveness is primary (Lindeberg, 2015, Lindeberg, 7 Oct 2025).

Another common misconception is that the time-causal limit kernel removes temporal delay. It does not. The papers emphasize that time-causal processing implies non-zero delay, with mean delay

$t<0$08

and an approximate mode

$t<0$09

What the logarithmic cascade changes is not the existence of delay, but the temporal dynamics: compared with uniform cascades, it shortens detection latency for a given composed variance (Lindeberg, 2015, Lindeberg, 7 Oct 2025).

Under illumination changes, the cited receptive-field theory further states that all non-zero order derivatives are invariant to additive brightness and, with log-brightness, multiplicative changes and exposure control variations, provided spatial derivatives are present; purely temporal derivatives $t<0$10 may respond to exposure fluctuations, which is one stated reason that spatial filtering in LGN is well justified (Lindeberg, 2015).

In this sense, the time-causal limit kernel occupies a specific position in scale-space theory: it is the canonical smoothing kernel for real-time temporal analysis when Gaussian smoothing is excluded by non-causal access to future data and when the smoothing process is required to be both variation diminishing and time recursive. Its importance lies less in closed-form simplicity than in the conjunction of properties it makes simultaneously available: strict causality, discrete scale covariance, recursive implementation, and a mathematically explicit relation between temporal scale, temporal memory, and feature responses (Lindeberg, 2022, Lindeberg, 2023).