Learned video codecs compress each frame by predicting it from the previous decoded frame and sending only a correction. Because the decoder only ever sees its own lossy output, small errors are fed back into every later prediction. Over hundreds of frames, quality sags, textures blur, and after a corrupted frame some codecs never recover. Reinforcement learning faced the same structure decades ago and solved it with one property. Value iteration, the repeated application of the Bellman update, tolerates per-step approximation error because that update is a contraction: it shrinks the distance between any two estimates by a factor $\gamma<1$. Old errors fade geometrically and only recent ones matter. This note proposes making the codec's closed decoding loop a contraction, so that it pulls itself back toward the true video.
The concrete claim is this. Treat the full loop of predicting, encoding the residual, quantizing and reconstructing as a map from "error in the reference frame" to "error in the next frame." Measure its local gain $\rho$ with two-frame perturbation probes, and train it so that $\rho<1$. Under that condition, the arithmetic behind approximate value iteration gives two results directly. First, drift has a horizon-independent ceiling, and recovery from corruption happens at a predictable rate. Second, bit allocation across frames becomes a well-posed backward recursion whose "discount factor" is the measured gain. A two-frame property certifies thousand-frame behaviour, which bears on the known gap that codecs are trained on 4–7 frame clips and then deployed on hundreds.
Core Claim
Let $\epsilon_t$ be the reconstruction error at frame $t$. Let $\rho_t$ be the gain of the closed-loop map $\epsilon_{t-1}\mapsto\epsilon_t$, with the encoder re-run so it can partly correct a perturbed reference. Let $\delta_t$ be the fresh error the loop would make from a perfect reference. Linearizing the recursion documented for learned codecs (Learned Video Compression for YUV 4:2:0 Content) gives
$$\|\epsilon_t\| \le \rho_t\,\|\epsilon_{t-1}\| + \delta_t ,$$
so if $\rho_t\le\rho<1$ the error approaches a ceiling of $\delta_{\max}/(1-\rho)$, and the influence of a corruption shrinks as $\rho^k$. This inequality has exactly the form of the approximate value iteration bound in Error Propagation for Approximate Policy and Value Iteration. That paper shows that because the Bellman update contracts, error injected at iteration $k$ is weighted by $\gamma^{K-k}$ ("later is greater"), and that shifting samples toward later iterations beats a uniform schedule. Here the codec's gain plays the role of $\gamma$, and forward time plays the role of iteration count.
The second half of the claim reverses the direction of the same analogy. Spending extra bits on frame $k$ improves frame $k$ and, through the loop, every later frame. The marginal value $w_k$ of quality at frame $k$, summed over the future, obeys
$$w_k = 1 + \rho_{k+1}\, w_{k+1},$$
which is a policy-evaluation recursion with a state-dependent discount $\rho_{k+1}$. It has a finite, unique solution on unbounded horizons only when $\rho<1$. When $\rho\ge 1$ the weights grow with sequence length. That is a precise account of why myopic and window-limited rate allocation fail on long sequences. In our view, closed-loop contraction is therefore the property that makes long-horizon rate control well posed, and it is more fundamental than either the codec architecture or the rate controller.
One distinction matters throughout. We do not ask the motion-compensation network alone to be contractive. A decoder that ignores its reference predicts badly, so that version would cost rate. The object we constrain includes the encoder's residual correction. Closed-loop predictive coding already has a corrective path, and the question is whether, at a given rate, that path removes more inherited error than the prediction amplifies.
Relation to Prior Work
Two papers motivate the gap by measuring drift without controlling the loop. EHVC shows that neural codecs relying on implicit dependencies degrade over long sequences and never recover after a gray frame is inserted. Its remedy is architectural: an explicit hierarchy that references a high-quality keyframe. Neural Video Compression with Feature Modulation (DCVC-FM) modulates temporal features to slow detail loss. Neither paper defines or bounds the loop's error gain. We reuse DCVC-FM as our base codec and EHVC's stress test as an evaluation, but we replace architectural intuition with a measured, trainable quantity.
The closest theoretical work is Causal State Compression for Long-Horizon Video World Models. It proves $\mathcal{O}(L^T\epsilon)$ drift for Lipschitz autoregressive rollouts and reduces it to $\mathcal{O}(\log^2 T\cdot\epsilon)$ with geometrically spaced keyframe memory. That paper constrains our design: it shows that an open-loop map with $L>1$ cannot be rescued without changing the dependency structure. We differ in three ways. Our lever is the closed loop, where a residual channel exists, so $L>1$ for prediction alone is compatible with $\rho<1$. Our setting is rate-constrained. And our bound is constant in $T$ rather than polylogarithmic, bought with a regularizer rather than a memory architecture.
Three rate-control papers operate on the same inter-frame dependency but never touch the loop's gain. Neural Rate Control for Learned Video Compression regresses per-frame importance weights inside miniGOP windows. Reinforced Rate Control treats bit allocation as a constrained Markov decision process and trains an actor-critic agent, with no analysis of whether its critic's update contracts. Our weight recursion supplies an analytic, measured counterpart to the first paper's learned weights. It also gives the second paper's critic a discount structure that comes from the codec rather than being chosen by hand.
On the reinforcement learning side, An Alternative Softmax Operator for Reinforcement Learning is the methodological precedent. It finds that a plausible operator (Boltzmann softmax) is not a non-expansion, shows it oscillating or having multiple fixed points on random MDPs, and repairs the property directly with mellowmax. We do the analogous thing for a codec: diagnose expansion, then enforce contraction with as small a change as possible.
Research Program
The first component is measurement. For a trained codec and a real sequence, we inject small perturbations into the decoded reference at frame $t-1$. We then re-run the full encoder–quantizer–decoder for frame $t$ and estimate $\rho_t$ as the largest amplification over perturbation directions. The estimate uses power iteration with Jacobian-vector products, with quantization handled by the uniform-noise relaxation introduced in DVC and checked against true rounding with finite differences.
We will report $\rho_t$ across content types, for several reasons. Scene cuts should make the reference irrelevant and $\rho_t$ small. High motion and low rate should push it up. The key diagnostic is whether DCVC-FM-class codecs sit above 1 in typical content. If they already sit below 1, drift is governed by the ceiling $\delta/(1-\rho)$, not by expansion, and the proposal narrows to its allocation half. We state this outcome in advance as a live possibility.
The second component is training. We add a hinge penalty $\max(0,\hat\rho_t-\rho^\star)^2$ to the usual rate–distortion loss. The penalty is computed on two-frame windows sampled from Vimeo-90k, and $\rho^\star$ is a target gain swept between 0.7 and 0.95. Because the penalty needs only one closed-loop step plus a few Jacobian-vector products, it adds no long backpropagation through time. That is the point: a local certificate standing in for long-horizon training.
The third component is allocation. At encode time, $\rho_t$ is estimated cheaply, either with a single power-iteration step or a small predictor distilled from it, and plugged into the weight recursion over a lookahead window. Because the recursion contracts, truncating it introduces error of order $\rho^{H}$ for a window of $H$ frames, so a short window suffices. Per-frame rate parameters are then set in proportion to $w_k$, under a sequence budget.
Baselines
The naive reference point is DCVC-FM itself, run as released on full-length sequences with a single intra frame and a fixed rate parameter. Every drift and recovery number is reported relative to it. This baseline answers whether anything is needed at all.
The strongest existing codec on long-sequence stability in the file is EHVC (EHVC). It reports an average BD-rate saving of 17.14% over VTM-23.4 across UVG, MCL-JCV and HEVC B–E, and degradation of about 1 dB versus about 1.5 dB for earlier models within a 32-frame cycle. Its hierarchical keyframe reference is the most direct competing answer to drift. We will use released EHVC results and, where code permits, rerun it on our exact sequence lengths and intra settings. Beating it means matching its BD-rate within 2% while showing a smaller end-to-start PSNR drop over 300 frames and a shorter gray-frame recovery time. We will also test whether our regularizer adds stability when applied on top of EHVC, since the two ideas are not mutually exclusive.
For rate control, the strongest baseline is Reinforced Rate Control, which reports an average relative bitrate error of 1.20% and up to 13.45% bitrate savings at typical GOP sizes. Next is Neural Rate Control, which reports about 2% bitrate error and nearly 10% BD-rate savings over fixed rate parameters. Both will be run on the same base codec, with the same budgets and test sets. Beating them means equal or lower bitrate error together with better BD-rate. The comparison should matter most on long GOPs, which is where the recursion predicts window-limited weights fail.
Two ablations isolate the key ingredient. The first applies the same regularizer to the motion-compensation network alone, open-loop, with matched tuning. If closed-loop structure is what matters, this variant should pay much more rate for the same drift reduction. The second keeps the contractive codec but allocates bits uniformly, or with Neural Rate Control's learned weights, which separates the two halves of the claim. All trained variants share the training data and GPU-hour budget, and get the same number of hyperparameter trials.
Why not just…?
Why not just train on longer sequences? This is the most serious alternative. Cascaded multi-frame training is how practitioners already fight drift, and with enough memory it may learn a contractive loop implicitly. We will train a DCVC-FM variant with rolled-out sequences of 16–32 frames, matched in GPU-hours to our two-frame regularized model, and compare drift at frame 300 and $\hat\rho_t$ distributions. If long training reaches the same $\rho$ and drift at equal compute, our contribution shrinks to a diagnostic plus an allocation rule. We would report it that way.
Why not just enforce a Lipschitz bound on the decoder, for example with spectral normalization? A global bound restricts every input direction, including those the prediction needs in order to follow motion. The Causal State Compression analysis implies that useful predictors often have $L>1$ in open loop. The open-loop ablation above tests exactly this. Our prediction is a substantially larger BD-rate penalty at matched drift.
Why not just refresh with periodic intra frames or a high-quality keyframe reference? Intra refresh resets error at a large rate cost. EHVC's keyframe reference is a softer version of the same move, and its reported 1 dB degradation within a 32-frame cycle shows drift is reduced, not removed. Refresh also bounds drift only between refreshes, and it does not help recovery from mid-cycle corruption. We compare against intra periods of 32 and 96 at matched rate, and expect the contractive loop to be complementary rather than a replacement.
Why not just let a reinforcement learning rate controller learn the inter-frame dependency? Reinforced Rate Control already does this empirically. However, its critic's target becomes horizon-dependent exactly when the loop expands, which is the regime where learning is hardest. If its learned allocation matches our analytic recursion on a contractive codec, that confirms the recursion is the right target, at much lower cost. If it beats the recursion, the gain estimates are missing something, and we will examine which content drives the difference.
Proposed Experiments
We will train on Vimeo-90k and evaluate on UVG, MCL-JCV and HEVC Classes B–E under low-delay settings, on full sequences up to 300+ frames with a single intra frame. Metrics are BD-rate against VTM in PSNR and MS-SSIM, the slope of per-frame PSNR against frame index, bitrate error under rate control, and the measured distribution of $\hat\rho_t$.
The decisive checks are four. First, whether two-frame $\hat\rho_t$ measured on the unregularized codec predicts its frame-300 drift across sequences, tested by rank correlation. If it does not, the linearized theory is wrong for these codecs and the program stops there. Second, whether the regularized codec's PSNR curve flattens to the predicted ceiling, at a BD-rate cost of under about 3%. Third, whether recovery after EHVC's gray-frame insertion follows the predicted half-life $\log 0.5/\log\rho$. Fourth, whether recursion-derived weights beat learned and RL allocations most clearly at long GOPs.
As a cross-domain check, the same gain measurement and penalty will be applied to an autoregressive model on BAIR Robot Pushing, where the closed loop is replaced by re-encoding of observed context. The reference point there is CSSC's reported 64-frame FVD of 96.4 against an autoregressive baseline of 218.7. The question is whether a contraction penalty alone recovers part of that gap without keyframe memory.
References
- Learned Video Compression for YUV 4: 2: 0 Content Using Flow-based Conditional Inter-frame Coding
- Error Propagation for Approximate Policy and Value Iteration
- EHVC: Efficient Hierarchical Reference and Quality Structure for Neural Video Coding
- Neural Video Compression with Feature Modulation
- Causal State Compression for Long-Horizon Video World Models: A Bounded-Drift Theory and Efficient Architecture
- Neural Rate Control for Learned Video Compression
- Reinforced Rate Control for Neural Video Compression via Inter-Frame Rate-Distortion Awareness
- An Alternative Softmax Operator for Reinforcement Learning
- DVC: An End-To-End Deep Video Compression Framework