Backward-Coherence Regularisation Overview

• Backward-Coherence Regularisation is a design pattern that enforces mutual compatibility between forward solutions and their associated backward operations.
• It integrates into the training objective—via inverse consistency, bidirectional temporal swaps, and hidden state reconstruction—to stabilize and improve model predictions.
• Applications range from deformable image registration and longitudinal MRI prediction to RNN dynamics and coherent inverse problems, with empirical improvements in error reduction and convergence.

to=arxiv_search.search 天天送json { "^^^^^^^^3query3 "^^^^^^^^3\3 coherence^^^^^^^^3\3 regularisation OR ^^^^^^^^3\3 "max_results": ^^^^3\3query3^^^^, "sort_by": "submittedDate" } to=arxiv_search.search 天天中彩票这个json { "^^^^^^^^3query3 "^^^^^^^^3\3 consistency^^^^^^^^3\3 registration regularization", "max_results": ^^^^3\3query3^^^^, "sort_by": "submittedDate" } Backward-Coherence Regularisation denotes a family of regularisation strategies in which model admissibility is constrained by consistency under a backward relation. In the available literature, that backward relation appears as inverse-consistent composition of spatial maps, bidirectional temporal prediction obtained by swapping baseline and follow-up scans, reconstructibility of an RNN hidden state from its successor, preservation of instantaneous phase while regularising only magnitude in coherent inverse problems, and coherence-enhancing constraints that stabilise backward reconstruction in diffusion MRI (&&&^^^^3query3^^^^&&&, &&&^^^^3\3^^^^&&&, &&&^^^^3 regularisation OR \3^^^^&&&, Watson et al., 2024, Hohage et al., 2014).

^^^^3\3^^^^. Scope and common structure

Across domains, backward-coherence regularisation is not a single canonical penalty but a design pattern. The regulariser couples a forward object to a backward map or backward direction: PRESERVED_PLACEHOLDER_^^^^3query3^^^^ to PRESERVED_PLACEHOLDER_^^^^3\3^^^^ in deformable registration, PRESERVED_PLACEHOLDER_^^^^3 regularisation OR \3^^^^ to $I_{T_b}$ and back in longitudinal MRI prediction, $h_t$ to $h_{t+1}$ in recurrent dynamics, and $z$ to $|z|$ with phase preserved in coherent imaging (&&&^^^^3query3^^^^&&&, &&&^^^^3\3^^^^&&&, &&&^^^^3 regularisation OR \3^^^^&&&, Watson et al., 2024). This suggests a unifying description: the regularisation acts by shrinking the feasible set to configurations that remain mutually compatible under inversion, reversal, or backward projection.

Setting	Backward relation	Regularisation form
Deformable registration	$\phi \circ \psi \approx Id$, $\psi \circ \phi \approx Id$	PRESERVED_PLACEHOLDER_^^^^3\3query3^^^^
Longitudinal diffusion MRI	baseline/follow-up swap, PRESERVED_PLACEHOLDER_^^^^3\3\3^^^^	BITR
RNN hidden states	PRESERVED_PLACEHOLDER_^^^^3\3 regularisation OR \3^^^^	PRESERVED_PLACEHOLDER_^^^^3\33^^^^
Coherent inverse problems	preserve PRESERVED_PLACEHOLDER_^^^^3\3^^^^4, regularise PRESERVED_PLACEHOLDER_^^^^3\3^^^^5	PRESERVED_PLACEHOLDER_^^^^3\3^^^^6

A recurrent feature is that backward coherence is embedded inside the optimisation or training objective rather than applied as post-processing. In registration, it is part of the loss; in longitudinal diffusion, it is a training-time direction swap; in RNNs, it augments the task objective; in coherent inverse problems, it changes the proximal step itself. The result is a regulariser whose semantics are task-specific but whose operational role is to suppress implausible forward solutions that cannot be reconciled with an associated backward operation.

^^^^3 regularisation OR \3^^^^. Inverse consistency in deformable registration

ICON defines backward-coherence as inverse consistency between a forward map PRESERVED_PLACEHOLDER_^^^^3\3^^^^7 and a backward map PRESERVED_PLACEHOLDER_^^^^3\3^^^^8, with PRESERVED_PLACEHOLDER_^^^^3\3^^^^9 and displacement parameterisation PRESERVED_PLACEHOLDER_^^^^3 regularisation OR \3query3^^^^, PRESERVED_PLACEHOLDER_^^^^3 regularisation OR \3\3^^^^ (&&&^^^^3query3^^^^&&&). The core loss is

PRESERVED_PLACEHOLDER_^^^^3 regularisation OR \3 regularisation OR \3^^^^

and, because ICON composes maps via interpolation, it is evaluated in displacement form as

PRESERVED_PLACEHOLDER_^^^^3 regularisation OR \33^^^^

This penalises departures from mutual invertibility: when PRESERVED_PLACEHOLDER_^^^^3 regularisation OR \3^^^^4, PRESERVED_PLACEHOLDER_^^^^3 regularisation OR \3^^^^5 and PRESERVED_PLACEHOLDER_^^^^3 regularisation OR \3^^^^6.

A crucial implementation detail is randomized off-grid evaluation. ICON samples points PRESERVED_PLACEHOLDER_^^^^3 regularisation OR \3^^^^7, with PRESERVED_PLACEHOLDER_^^^^3 regularisation OR \3^^^^8 drawn from a multivariate Gaussian whose per-dimension standard deviation equals the pixel or voxel spacing, and evaluates the inverse-consistency residual away from lattice centres. The stated rationale is that on-grid evaluation alone allows a “swap” of neighbouring grid points in a textureless region to compose to the identity at those centres while still folding between them. Off-grid sampling exposes this non-invertibility within cells and raises PRESERVED_PLACEHOLDER_^^^^3 regularisation OR \3^^^^9.

ICON also supplies a theoretical intuition for why inverse consistency enforces spatial regularity. Under small, independent spatial white-noise perturbations of the output maps before composition, the expected inverse-consistency term expands into the original coherence loss plus terms proportional to

$I_{T_b}$^^^^3query3^^^^

The paper interprets this as an implicit $I_{T_b}$^^^^3\3^^^^-type regularisation weighted by local volume change. Even without explicit injected noise, ICON argues that training error in inverse consistency acts similarly because finite network capacity and stochastic training prevent exact satisfaction of the inverse-consistency equations across a population.

The full training objective is

$I_{T_b}$^^^^3 regularisation OR \3^^^^

ICON does not add tuned smoothness, elastic, diffusion, or bending-energy penalties such as $I_{T_b}$3. Continuity and differentiability follow from linear interpolation except on a measure-zero set, approximate invertibility is enforced by $I_{T_b}$4, and folding is quantified via the fraction of voxels where $I_{T_b}$5.

Empirically, inverse consistency plus off-grid interpolation reduces folds compared to similarity alone. On synthetic shapes, U-Net achieved Dice $I_{T_b}$6–$I_{T_b}$7 with folds reduced to $I_{T_b}$8–$I_{T_b}$9 as $h_t$^^^^3query3^^^^ grows; on the 3D OAI knee dataset, ICON (^^^^3 regularisation OR \3^^^^ half-res + ^^^^3 regularisation OR \3^^^^ full-res) reported Dice $h_t$^^^^3\3^^^^, folds $h_t$^^^^3 regularisation OR \3^^^^, and time $h_t$3 s, while SyN (CC) achieved Dice $h_t$4 with $h_t$5 folds but required $h_t$6 s. ICON therefore presents backward coherence as a standalone regulariser for approximately diffeomorphic maps, while also stating that backward-coherence alone does not strictly guarantee diffeomorphisms at test time.

3. Bidirectional temporal regularisation in longitudinal diffusion models

In TADM-3D, backward-coherence regularisation appears as Back-In-Time Regularisation (BITR), a training strategy for a 3D Denoising Diffusion Probabilistic Model that predicts a future MRI from a baseline volume and a desired time interval $h_t$7 (&&&^^^^3\3^^^^&&&). The model predicts the voxel-wise intensity residual

$h_t$8

and reconstructs the follow-up through

$h_t$9

Temporal conditioning uses a baseline latent $h_{t+1}$^^^^3query3^^^^, the interval $h_{t+1}$^^^^3\3^^^^, age at baseline $h_{t+1}$^^^^3 regularisation OR \3^^^^, and cognitive status $h_{t+1}$3.

BITR is implemented by randomly swapping $h_{t+1}$4 with probability $h_{t+1}$5. In the forward direction, the model predicts $h_{t+1}$6 from $h_{t+1}$7 using $h_{t+1}$8; in the backward direction, it predicts $h_{t+1}$9 from $z$^^^^3query3^^^^ using $z$^^^^3\3^^^^. The paper states that predicting past scans has limited clinical applications, but that this regularisation helps the model generate temporally more accurate scans.

The training objective combines a diffusion $z$^^^^3 regularisation OR \3^^^^-prediction term with an age-consistency penalty from a pre-trained Brain-Age Estimator: $z$3

$z$4

$z$5

The paper explicitly states that there is no separate $z$6 and $z$7; BITR is a training-time data-direction augmentation. It also states that it does not add a cycle-consistency loss.

The empirical role of backward coherence is pronounced in the ablations. On the internal OASIS-3 test set, full TADM-3D achieved MSE $z$8 and SSIM $z$9, whereas the variant w/o BITR achieved MSE $|z|$^^^^3query3^^^^ and SSIM $|z|$^^^^3\3^^^^. The paper summarises this as a substantial degradation when BITR is removed: MSE doubles and SSIM drops. On the external NACC test set, TADM-3D achieved MSE $|z|$^^^^3 regularisation OR \3^^^^ and SSIM $|z|$3, with best regional MAE on hippocampus, amygdala, thalamus, and CSF, while BrLP was lowest on lateral ventricle.

The model therefore uses backward coherence not as explicit inversion but as temporal bidirectionality. The stated theoretical intuition is that learning both positive and negative residuals reduces drift and biases, improving sensitivity to the sign and magnitude of $|z|$4. A related practical distinction is that conditioning is on age difference rather than absolute age, and that a frozen BAE provides a direct temporal scalar constraint rather than a cycle penalty.

4. Hidden-state stability in recurrent neural networks

In recurrent models, backward-coherence regularisation is formulated as probabilistic backward consistency of hidden states: a stable hidden state should be reconstructible from its successor (&&&^^^^3 regularisation OR \3^^^^&&&). For an RNN with hidden state $|z|$5 and update $|z|$6, the backward filtration is

$|z|$7

Backward coherence requires

$|z|$8

with a learned backward projector $|z|$9 in residual form,

$\phi \circ \psi \approx Id$^^^^3query3^^^^

The observable backward residual and reverse-martingale drift surrogate are

$\phi \circ \psi \approx Id$^^^^3\3^^^^

with empirical totals

$\phi \circ \psi \approx Id$^^^^3 regularisation OR \3^^^^

The backward-coherence penalty is

$\phi \circ \psi \approx Id$3

and the total objective is

$\phi \circ \psi \approx Id$4

The theoretical analysis proceeds under contraction in the hidden-state argument,

$\phi \circ \psi \approx Id$5

backward sufficiency,

$\phi \circ \psi \approx Id$6

and summable backward drift. Under these assumptions, the hidden-state sequence forms a reverse quasi-martingale and Theorem 3.^^^^3\3^^^^ gives almost-sure convergence when $\phi \circ \psi \approx Id$7. The paper also gives $\phi \circ \psi \approx Id$8 convergence under uniform integrability, geometric convergence under exponentially decaying drift defects, change-point tracking bounds, and time-uniform confidence sequences based on increment-sum tubes and a calibrated defect-tail proxy.

Backward coherence also receives a variational interpretation. With

$\phi \circ \psi \approx Id$9

minimising $\psi \circ \phi \approx Id$^^^^3query3^^^^ over $\psi \circ \phi \approx Id$^^^^3\3^^^^ is equivalent to minimising an expected Kullback–Leibler divergence to a Gaussian backward model with the correct conditional mean. The paper further derives the bound

$\psi \circ \phi \approx Id$^^^^3 regularisation OR \3^^^^

which links the scalar diagnostic $\psi \circ \phi \approx Id$3 to the training loss.

Empirically, backward-coherence regularisation reduces the empirical quasi-martingale total $\psi \circ \phi \approx Id$4 by $\psi \circ \phi \approx Id$5–$\psi \circ \phi \approx Id$6 and reaches stability $\psi \circ \phi \approx Id$7–$\psi \circ \phi \approx Id$8 earlier than an unregularised RNN. On PhysioNet ^^^^3 regularisation OR \3query3\3 regularisation OR \3^^^^ ICU data, the Reverse Martingale RNN matched mortality-prediction AUC, $\psi \circ \phi \approx Id$9 versus PRESERVED_PLACEHOLDER_^^^^3\3query3query3^^^^ baseline, while reaching stable representations PRESERVED_PLACEHOLDER_^^^^3\3query3\3^^^^ hours earlier. On FRED-MD, it reduced one-month-ahead forecast error by about fourfold under concept drift. On UCI Human Activity Recognition, it maintained lower post-transition tracking error with geometric decay. The paper simultaneously emphasises that the guarantees apply under the stated assumptions and that universality is not claimed.

5. Magnitude-only proximal regularisation in coherent inverse problems

In coherent inverse problems such as SAR, backward-coherence regularisation has a different technical meaning: regularise only the magnitudes and preserve the instantaneous phase of the current iterate (Watson et al., 2024). The forward model is PRESERVED_PLACEHOLDER_^^^^3\3query3 regularisation OR \3^^^^ with PRESERVED_PLACEHOLDER_^^^^3\3query33^^^^, and the image is written componentwise as

PRESERVED_PLACEHOLDER_^^^^3\3query3^^^^4

The prior information is assumed to reside primarily in PRESERVED_PLACEHOLDER_^^^^3\3query3^^^^5, while phase is often treated as weakly constrained.

For a proper, closed, convex PRESERVED_PLACEHOLDER_^^^^3\3query3^^^^6 on magnitudes, define

PRESERVED_PLACEHOLDER_^^^^3\3query3^^^^7

If PRESERVED_PLACEHOLDER_^^^^3\3query3^^^^8, then the main theorem gives the phase-preserving proximal map

PRESERVED_PLACEHOLDER_^^^^3\3query3^^^^9

The paper’s interpretation is that the magnitude penalty acts “backward” with respect to phase: phase is updated solely by the data-fidelity step, while the regulariser corrects only magnitudes.

When the sufficient condition fails, the paper introduces the bounded proximal map

PRESERVED_PLACEHOLDER_^^^^3\3\3query3^^^^

and the correct complex proximal becomes

PRESERVED_PLACEHOLDER_^^^^3\3\3\3^^^^

PRESERVED_PLACEHOLDER_^^^^3\3\3 regularisation OR \3^^^^ is computed through Douglas–Rachford splitting,

PRESERVED_PLACEHOLDER_^^^^3\3\33^^^^

with

PRESERVED_PLACEHOLDER_^^^^3\3\3^^^^4

This framework is instantiated for magnitude sparsity, generalised Tikhonov, TGV, and level-set regularisation. For PRESERVED_PLACEHOLDER_^^^^3\3\3^^^^5, the induced complex proximal is

PRESERVED_PLACEHOLDER_^^^^3\3\3^^^^6

For PRESERVED_PLACEHOLDER_^^^^3\3\3^^^^7, the direct formula

PRESERVED_PLACEHOLDER_^^^^3\3\3^^^^8

holds under the stated sufficient conditions; otherwise the Douglas–Rachford correction is used.

The paper demonstrates the method on publicly available real SAR data for generalised Tikhonov regularisation applied to multi-channel SAR, and both a simple level set formulation and total generalised variation applied to the standard single-channel case. The stated benefits are that no nonlinear phase fitting is needed, regularisers do not fight the data term by trying to homogenise phases, and interferometric uses are preserved while magnitude structure is improved. The main limitation is equally explicit: in extreme noise or severe model mismatch, the data term may drive phases erratically, and backward-coherence regularisation will not correct phase ambiguities.

6. Antecedents, adjacent formulations, and related coherence principles

Several adjacent formulations illuminate the broader landscape in which backward coherence is used. In diffusion MRI, the paper on a coherence enhancing penalty reconstructs orientation distribution functions PRESERVED_PLACEHOLDER_^^^^3\3\3^^^^9 from HARDI data by solving a linear inverse problem regularised by fiber continuity (Hohage et al., 2014). The geometric assumption is

PRESERVED_PLACEHOLDER_^^^^3\3 regularisation OR \3query3^^^^

which yields

PRESERVED_PLACEHOLDER_^^^^3\3 regularisation OR \3\3^^^^

and the coherence-enhancing penalty

PRESERVED_PLACEHOLDER_^^^^3\3 regularisation OR \3 regularisation OR \3^^^^

The full constrained objective is

PRESERVED_PLACEHOLDER_^^^^3\3 regularisation OR \33^^^^

subject to PRESERVED_PLACEHOLDER_^^^^3\3 regularisation OR \3^^^^4. The paper states that this stabilises the backward reconstruction by enforcing low variance along PRESERVED_PLACEHOLDER_^^^^3\3 regularisation OR \3^^^^5-directions across voxels and proves convergence for discrete, noisy data through compact embedding of PRESERVED_PLACEHOLDER_^^^^3\3 regularisation OR \3^^^^6 into PRESERVED_PLACEHOLDER_^^^^3\3 regularisation OR \3^^^^7 and constrained Tikhonov theory with operator approximation.

A different but mathematically related use of backward induction appears in sequential coherence for predictive modelling (&&&^^^^3\3^^^^4&&&). There, predictive densities satisfy

PRESERVED_PLACEHOLDER_^^^^3\3 regularisation OR \3^^^^8

or equivalently

PRESERVED_PLACEHOLDER_^^^^3\3 regularisation OR \3^^^^9

so PRESERVED_PLACEHOLDER_^^^^3\33query3^^^^ is a martingale. Starting from a terminal predictive PRESERVED_PLACEHOLDER_^^^^3\33\3^^^^, backward induction yields a time-consistent predictive sequence and, in the kernel density example, produces earlier densities as scale mixtures with stochastic bandwidth inflation. The paper describes this as coherent, prior-free uncertainty assessment and derives Azuma–Hoeffding concentration through bounded one-step predictive differences.

Backward coherence also appears in optimisation-theoretic form through backward error analysis of SGD and differentiable games (&&&^^^^3\3^^^^5&&&). For one-step GD, the modified flow is

PRESERVED_PLACEHOLDER_^^^^3\33 regularisation OR \3^^^^

For multiple consecutive SGD steps on exact batches PRESERVED_PLACEHOLDER_^^^^3\333^^^^, the paper constructs an iteration-dependent modified loss with a full-batch gradient-norm penalty and an inter-batch gradient alignment term depending explicitly on PRESERVED_PLACEHOLDER_^^^^3\33^^^^4. In two-player differentiable games it similarly obtains per-iteration scalar modified losses containing self-gradient penalties and cross-player interaction terms frozen at the previous iterate. The common device is to exploit the BEA freedom to choose correction fields that depend on the initial point, thereby making the modified vector field integrable.

Taken together, these formulations suggest that backward coherence is best understood as a recurring regularisation principle rather than a single algorithm. The backward object may be an inverse deformation, a reversed temporal pair, a successor hidden state, a retained phase, a terminal predictive distribution, or a previous iterate in a modified flow. What remains stable across the literature is the role of the regulariser: it penalises forward solutions that are incompatible with a specified backward relation, and thereby replaces or supplements conventional smoothness, isotropy, or prior-based constraints.