Temporal Transfer Matrix Overview

• Temporal Transfer Matrix is a representation that converts a univariate time series into a T×T image using quantile-based states and local Markov transition probabilities.
• It segments the series into K contiguous chunks to estimate local transition matrices, thereby revealing nonstationary regime changes while balancing bias and variance.
• The method is amplitude-agnostic and order-preserving, making it effective for CNN-based classification, regime detection, and anomaly analysis in time series.

Searching arXiv for recent and related uses of “Temporal Transfer Matrix” and “Temporal Markov Transition Field”. Temporal Transfer Matrix, in the sense of the Temporal Markov Transition Field (TMTF), is a temporally aware $T\times T$ matrix representation of a univariate time series in which “transfer” and “transition” are synonymous: both denote matrices of conditional probabilities of moving between quantile-defined states. The construction extends the Markov Transition Field by partitioning the series into $K$ contiguous temporal chunks, estimating a separate local transition matrix for each chunk, and assigning each row according to the dynamics active in that row’s chunk. The resulting image exposes when regime changes occur, while remaining amplitude-agnostic and order-preserving, and is suitable as an input channel for convolutional neural networks applied to time series characterisation tasks (Leznik, 9 Mar 2026).

1. Quantile encoding and the global Markov Transition Field

Let the observed series be $x_1,x_2,\ldots,x_T$. Choose $Q\in\mathbb{N}$ quantile bins, with empirical quantile boundaries $q_0<q_1<\cdots<q_Q$ such that each interval $[q_{k-1},q_k)$ contains approximately $\lfloor T/Q\rfloor$ observations, adjusting at boundaries if needed. The induced state sequence is

$$s_t = k \quad \text{if } x_t \in [q_{k-1},q_k), \qquad k=1,\ldots,Q.$$

This encoding is amplitude-agnostic and order-preserving: any strictly increasing transformation $f$ applied to $\{x_t\}$ preserves ranks, thus $K$0 and $K$1 are unchanged. Increasing $K$2 increases resolution, but requires more data to reliably estimate transitions (Leznik, 9 Mar 2026).

The global Markov Transition Field assumes stationary transition dynamics across time. Its first-order Markov transition matrix is

$K$3

with each row summing to $K$4. The corresponding image $K$5 is defined by

$K$6

This global construction is efficient when transition dynamics are stationary, but it has a structural limitation. Because all transitions are pooled into a single $K$7, the image has at most $K$8 distinct row patterns: if $K$9, then row $x_1,x_2,\ldots,x_T$0 equals row $x_1,x_2,\ldots,x_T$1. The consequence is row degeneracy: the global MTF cannot reveal when dynamics change, because all time steps in the same state share the same row regardless of temporal location (Leznik, 9 Mar 2026).

2. Temporal chunking and the TMTF construction

The TMTF replaces the single global transition matrix with chunk-specific local matrices. Time indices are split into $x_1,x_2,\ldots,x_T$2 contiguous, non-overlapping chunks $x_1,x_2,\ldots,x_T$3 covering $x_1,x_2,\ldots,x_T$4. For equal-length chunks, when $x_1,x_2,\ldots,x_T$5,

$x_1,x_2,\ldots,x_T$6

with chunk index function $x_1,x_2,\ldots,x_T$7 such that $x_1,x_2,\ldots,x_T$8 (Leznik, 9 Mar 2026).

For each chunk $x_1,x_2,\ldots,x_T$9, a local transition matrix is estimated using only within-chunk consecutive transitions and excluding pairs that cross chunk boundaries: $Q\in\mathbb{N}$0

The temporally aware image is then

$Q\in\mathbb{N}$1

Only the row index is chunked: the row is governed by the local dynamics of the row’s chunk, while the column index remains the destination state $Q\in\mathbb{N}$2 at time $Q\in\mathbb{N}$3. This column asymmetry preserves cross-chunk comparability in columns: differences across horizontal bands reflect changing departure dynamics, not differing destinations (Leznik, 9 Mar 2026).

The image therefore exhibits $Q\in\mathbb{N}$4 horizontal bands. Rows $Q\in\mathbb{N}$5 share the same governing local matrix $Q\in\mathbb{N}$6, and textures differ across bands if the $Q\in\mathbb{N}$7 vary by segment. Rows within a band remain grouped by state $Q\in\mathbb{N}$8, but TMTF can show up to $Q\in\mathbb{N}$9 distinct row patterns, rather than the $q_0<q_1<\cdots<q_Q$0 patterns of the global MTF (Leznik, 9 Mar 2026).

3. Structural properties and statistical trade-offs

TMTF preserves several invariances and reductions inherited from quantile encoding. It is amplitude-agnostic and order-preserving, because it is invariant under strictly increasing transforms $q_0<q_1<\cdots<q_Q$1 of $q_0<q_1<\cdots<q_Q$2. It encodes first-order Markov structure, namely $q_0<q_1<\cdots<q_Q$3, not higher-order dependencies. It also degrades gracefully: if $q_0<q_1<\cdots<q_Q$4, then $q_0<q_1<\cdots<q_Q$5, so the temporal construction reduces exactly to the global MTF in stationary settings (Leznik, 9 Mar 2026).

Temporal chunking introduces an explicit bias–variance trade-off. Variance rises because each $q_0<q_1<\cdots<q_Q$6 is estimated from fewer transitions, approximately $q_0<q_1<\cdots<q_Q$7, with rough per-entry variance scaling

$q_0<q_1<\cdots<q_Q$8

where $q_0<q_1<\cdots<q_Q$9 is the empirical frequency of state $[q_{k-1},q_k)$0 in chunk $[q_{k-1},q_k)$1. Bias falls when the true dynamics vary by chunk: the global $[q_{k-1},q_k)$2 is biased for each segment, whereas the local $[q_{k-1},q_k)$3 is unbiased within its chunk and reduces bias compared to pooling (Leznik, 9 Mar 2026).

A practical rule is to target at least $[q_{k-1},q_k)$4 transitions per row on average, implying

$[q_{k-1},q_k)$5

For $[q_{k-1},q_k)$6 and $[q_{k-1},q_k)$7, $[q_{k-1},q_k)$8 is reported as a robust default, for example when $[q_{k-1},q_k)$9 (Leznik, 9 Mar 2026).

The computational profile is straightforward. Quantile binning is $\lfloor T/Q\rfloor$0 if sorting is required, state sequence construction is $\lfloor T/Q\rfloor$1, estimating all $\lfloor T/Q\rfloor$2 local matrices is $\lfloor T/Q\rfloor$3, assembling $\lfloor T/Q\rfloor$4 is $\lfloor T/Q\rfloor$5 lookups, and memory is $\lfloor T/Q\rfloor$6. For very long series, downsampling, tiling, or low-rank approximations are suggested (Leznik, 9 Mar 2026).

4. Geometric interpretation and canonical example

The local matrices admit a direct geometric interpretation in terms of process properties. Large diagonal entries $\lfloor T/Q\rfloor$7 indicate persistence: the process tends to remain in its current quantile band, yielding a dark main diagonal stripe within the corresponding band. Small diagonal mass with substantial off-diagonal mass indicates mean reversion and produces diffuse texture. Upper-triangular emphasis indicates trending upward, while a nearly uniform matrix corresponds to near-random behaviour in which the current state carries little information about the next (Leznik, 9 Mar 2026).

The paper’s worked example makes the contrast with the global MTF explicit. For

$\lfloor T/Q\rfloor$8

with $\lfloor T/Q\rfloor$9, $$s_t = k \quad \text{if } x_t \in [q_{k-1},q_k), \qquad k=1,\ldots,Q.$$0, and $$s_t = k \quad \text{if } x_t \in [q_{k-1},q_k), \qquad k=1,\ldots,Q.$$1, the quantile state sequence is

$$s_t = k \quad \text{if } x_t \in [q_{k-1},q_k), \qquad k=1,\ldots,Q.$$2

The global tally over $$s_t = k \quad \text{if } x_t \in [q_{k-1},q_k), \qquad k=1,\ldots,Q.$$3 consecutive transitions gives

$$s_t = k \quad \text{if } x_t \in [q_{k-1},q_k), \qquad k=1,\ldots,Q.$$4

which averages the mean-reverting first half and the persistent or trending second half, thereby obscuring when the regimes occurred (Leznik, 9 Mar 2026).

For TMTF with $$s_t = k \quad \text{if } x_t \in [q_{k-1},q_k), \qquad k=1,\ldots,Q.$$5 and $$s_t = k \quad \text{if } x_t \in [q_{k-1},q_k), \qquad k=1,\ldots,Q.$$6, the local matrices are:

Chunk	Local transition matrix	Interpretation
$$s_t = k \quad \text{if } x_t \in [q_{k-1},q_k), \qquad k=1,\ldots,Q.$$7	$$s_t = k \quad \text{if } x_t \in [q_{k-1},q_k), \qquad k=1,\ldots,Q.$$8	Zero diagonal; deterministic cycle; strong mean-reversion texture
$$s_t = k \quad \text{if } x_t \in [q_{k-1},q_k), \qquad k=1,\ldots,Q.$$9	$f$0	Large diagonal entries; strictly upper-triangular; state 3 absorbing

The TMTF image

$f$1

therefore shows a top band with crisp off-diagonal patterns and a bottom band with heavy diagonal and upper-triangular dominance. The horizontal texture change at $f$2 clearly reveals when the regime switched. The paper further notes a generative interpretation: in AR(1) terms, chunk 1 resembles $f$3, while chunk 2 resembles $f$4 with upward bias (Leznik, 9 Mar 2026).

5. Practical use, applications, and limitations

Practical design revolves mainly around choosing $f$5 and $f$6. The paper states that $f$7 works well for $f$8 in practice. Larger $f$9 yields finer distinctions but higher variance. Multi-resolution use is suggested: multiple TMTF channels at different $\{x_t\}$0 can balance robustness and discrimination. For $\{x_t\}$1, domain knowledge about possible regime durations is recommended, together with the minimum-transition rule $\{x_t\}$2. Cross-validation or model selection can evaluate downstream performance as a function of $\{x_t\}$3, and multi-scale chunking can combine several $\{x_t\}$4, for example $\{x_t\}$5, as separate channels (Leznik, 9 Mar 2026).

Regularization is important when chunks are small. Laplace, or add-$\{x_t\}$6, smoothing is proposed: $\{x_t\}$7 Shrinkage toward the global $\{x_t\}$8 or toward a prior such as the uniform distribution is also suggested (Leznik, 9 Mar 2026).

Applications are stated in operational rather than theoretical terms. $\{x_t\}$9 can serve as a CNN input channel, and multiple channels may be stacked, for example several $K$00, several $K$01, or combinations with raw or normalized series. Listed tasks include classification, regime detection or change point analysis, stationarity characterization, and anomaly detection. Because the representation is rank-based, heavy amplitude normalization is unnecessary, although optional per-image scaling may be used for numerical stability (Leznik, 9 Mar 2026).

The main limitations are sensitivity to $K$02 and $K$03, chunk boundary effects, and the restriction to first-order dynamics. Transitions crossing chunk boundaries are excluded, so overlapping chunks or soft chunking are proposed to reduce boundary artifacts. Soft assignments through kernelized ranks are suggested to reduce discretization artifacts at quantile boundaries. Extensions mentioned in the paper include online or streaming updates, multivariate encodings by channel stacking or joint-state quantiles, higher-order Markov fields, and variable-width bins (Leznik, 9 Mar 2026).

6. Terminological scope and cross-disciplinary usage

Within the TMTF literature, “Temporal Transfer Matrix” means a matrix of conditional probabilities that transfers mass from current to next quantile-defined state, with a separate transfer matrix $K$04 estimated for each temporal segment. In that precise sense, the “Temporal Transfer Matrix” is the $K$05 TMTF image whose entries transfer probabilities from $K$06 to $K$07 under the dynamics active at time $K$08 (Leznik, 9 Mar 2026).

The same phrase, however, appears in several unrelated technical traditions. In optical scattering, a temporal transmission matrix is a time-resolved mapping

$K$09

defined at discrete time bins under time-gated measurement (Guo et al., 2020). In one-dimensional scattering theory, the “temporal transfer matrix” is the truncation-dependent operator $K$10 satisfying

$K$11

with $K$12 playing the role of an evolution parameter (Mostafazadeh, 2020). In time-varying electromagnetic media, one finds a temporal ABCD matrix acting on $K$13 rather than $K$14 (Molero et al., 19 Mar 2025), as well as temporal multilayer transfer matrices for higher-order transfer functions (Ramaccia et al., 5 Feb 2025), generalized transfer matrices for space-time modulation (Li et al., 2019), temporal T-matrix formalisms for broadband or moving scatterers (Vavilin et al., 2023), and time-modulated acoustic-waveguide transfer matrices (Cidlinsky et al., 4 Apr 2025). The label also appears in exceptional-point media (Wang et al., 3 Nov 2025), integrable Floquet circuits (Vilkoviskiy et al., 19 Aug 2025), causal dynamical triangulations (Ambjorn et al., 2012), and tensor-network treatments of quantum transfer matrices and imaginary-time propagation (Zauner et al., 2014).

This suggests that “Temporal Transfer Matrix” is not a field-invariant object but a domain-dependent designation. In the present sense, it denotes the Temporal Markov Transition Field: a chunk-local, order-preserving, amplitude-agnostic image representation for nonstationary time series that restores the missing temporal localisation of regime-dependent transition dynamics (Leznik, 9 Mar 2026).