Literature
Display
← All papers

Preserving Unstable Modes Through Inverse Dynamics in JEPA World Models

Leonardo F. Toso · Yann LeCun · James Anderson · Oumayma Bounou

· arXiv (preprint) ↗

World models7 min text · CondensedOriginal paper ↗
Contents

Overview

The official project and code accompany the paper. The reading separates its exact linear-system conditions from empirical nonlinear control results.

At a glance

Intuition. A model can predict the background perfectly yet fail to notice a pole beginning to fall. Avoiding a constant embedding does not guarantee that the representation preserves the unstable directions a controller must observe. This paper constructs linear examples where next-state prediction and Gaussian anti-collapse are both optimized while those directions disappear. Its remedy is endpoint inverse dynamics: recover the applied action sequence from initial and final latents, encouraging action-reachable information to survive. Exact recovery implies a strong preservation result under linearity, rich action excitation and dimension/rank conditions. In visual CartPole and Walker2D experiments, the extra loss improves feedback-relevant dynamics. The nonlinear results are empirical, use the specified observations including proprioception, and do not extend the exact linear theorem into a blanket stabilization guarantee.

1 Introduction

Feedback control needs information about directions that grow without correction. A representation may remain globally varied and predictably evolve while hiding precisely those directions. The paper therefore evaluates control-aware preservation separately from anti-collapse and prediction quality, using both linear analysis and nonlinear visual tasks.

Refer to caption
Figure 1: Architecture of our proposed world model. At each step, an encoder (EE) maps observation yty_{t} to a latent state ztz_{t}, and a predictor (PP) rolls out latent predictions z^t+1,…,z^t+H\hat{z}_{t+1},\ldots,\hat{z}_{t+H} conditioned on actions. Training losses are applied at each predicted latent: 1SP/MSP (single- or multi-step prediction) aligns z^t+k\hat{z}_{t+k} with the encoded ground-truth zt+kz_{t+k}. The EP-IDM estimates the intermediate actions 𝐚^t,H=(a^t,a^t+1,…,a^t+H−1)\hat{\mathbf{a}}_{t,H}=(\hat{a}_{t},\hat{a}_{t+1},\ldots,\hat{a}_{t+H-1}) from ztz_{t} and zt+Hz_{t+H}.

2 Related Work

The work connects latent JEPA prediction with detectability and stabilization. Prior inverse-dynamics objectives already used action information; this endpoint version reconstructs an entire action block and analyzes which reachable state directions that requirement preserves.

3 Setup

The physical system, observation encoder and latent predictor are separate objects. Observations can contain images and proprioception; experiments use motion information from consecutive frames plus the stated state measurements. Standard prediction and marginal anti-collapse do not specify which physical variables must be recoverable.

Equations 1–7
xt+1=f⁡(xt,at),yt=g⁡(xt)∈𝒴⊂ℝp,x_{t+1}=f(x_{t},a_{t}),\qquad y_{t}=g(x_{t})\in\mathcal{Y}\subset\mathbb{R}^{p}, (1)
zt+1=Pθ​(zt,Gθ​(at))​ with latent state ​zt=Eθ​(yt),z_{t+1}=P_{\theta}\bigl(z_{t},G_{\theta}(a_{t})\bigr)\text{ with latent state }z_{t}=E_{\theta}(y_{t}), (2)
τ:=(ytτ−1:tτ+H,stτ:tτ+H,atτ:tτ+H−1),\displaystyle\tau:=\left(y_{t_{\tau}-1:t_{\tau}+H},s_{t_{\tau}:t_{\tau}+H},a_{t_{\tau}:t_{\tau}+H-1}\right), (3)
ztτ+k=Eθ(ytτ+k−1,ytτ+k,ytτ+k−ytτ+k−1,stτ+k),k=0,…,H.\displaystyle z_{t_{\tau}+k}=E_{\theta}\left(y_{t_{\tau}+k-1},y_{t_{\tau}+k},y_{t_{\tau}+k}-y_{t_{\tau}+k-1},s_{t_{\tau}+k}\right),\qquad k=0,\ldots,H. (4)
ℒpred​(θ)=𝔼τ∼𝒟​[1H​∑k=0H−1‖Pθ​(ztτ+k,Gθ​(atτ+k))−ztτ+k+1‖22].\mathcal{L}_{\mathrm{pred}}(\theta)=\mathbb{E}_{\tau\sim\mathcal{D}}\left[\frac{1}{H}\sum_{k=0}^{H-1}\left\|P_{\theta}\bigl(z_{t_{\tau}+k},G_{\theta}(a_{t_{\tau}+k})\bigr)-z_{t_{\tau}+k+1}\right\|_{2}^{2}\right]. (5)
minθ⁡ℒpred​(θ)+λreg​ℒreg​(θ),\min_{\theta}\;\mathcal{L}_{\mathrm{pred}}(\theta)+\lambda_{\mathrm{reg}}\mathcal{L}_{\mathrm{reg}}(\theta), (6)
min𝒂\displaystyle\min_{\boldsymbol{a}} ‖zT−zg‖22subject to\displaystyle\|z_{T}-z_{\mathrm{g}}\|_{2}^{2}\qquad\text{subject to}\quad zt+1=Pθ(zt,Gθ(at)),t=0,…,T−1,\displaystyle z_{t+1}=P_{\theta}\bigl(z_{t},G_{\theta}(a_{t})\bigr),\qquad t=0,\ldots,T-1, (7)

4 Prediction Does Not Guarantee Stabilization

A latent controller may stabilize its own predicted state while the true physical state diverges in an omitted unstable direction. The linear setup makes this failure precise under stabilizable physical dynamics and observable raw measurements, separating insufficient learned representation from an inherently uncontrollable system.

Equations 8–10
xt+1\displaystyle x_{t+1} =A​xt+B​at,yt=C​xt,\displaystyle=Ax_{t}+Ba_{t},\qquad y_{t}=Cx_{t}, (8)
rank⁡[λ​I−AM]=nfor every ​λ∈spec⁡(A)​ such that ​|λ|≥1,\displaystyle\operatorname{rank}\begin{bmatrix}\lambda I-A\\ M\end{bmatrix}=n\quad\text{for every }\lambda\in\operatorname{spec}(A)\text{ such that }|\lambda|\geq 1, (9)
Eθ​(y)=W​y+b=z∈𝒵,Pθ​(z,a)=Az​z+Bz​a∈𝒵.\displaystyle E_{\theta}(y)=Wy+b=z\in\mathcal{Z},\qquad P_{\theta}(z,a)=A_{z}z+B_{z}a\in\mathcal{Z}. (10)

4.1 Necessary and Sufficient Conditions for Stabilization of LDS

A controller using the encoded history can stabilize the analyzed linear system exactly when the encoded observation is detectable: every invisible mode must already decay without feedback. The criterion includes marginally stable as well as unstable modes. It concerns available information, not the performance of an arbitrary learned controller.

Equation 11
A​v=λ​v,F​v=0,|λ|≥1.\displaystyle Av=\lambda v,\qquad Fv=0,\qquad|\lambda|\geq 1. (11)

4.2 The Collapse of Unstable Modes

The counterexample discards unstable coordinates and whitens a predictable stable component. It can achieve zero one-step error and the desired Gaussian latent law while losing the information required for stabilization. Thus neither a rich marginal distribution nor perfect latent prediction alone certifies control relevance.

Equations 12–13
A=[AuΔ0As],B=[BuBs], and ​xt=[xu,txs,t].\displaystyle A=\begin{bmatrix}A_{u}&\Delta\\ 0&A_{s}\end{bmatrix},\;B=\begin{bmatrix}B_{u}\\ B_{s}\end{bmatrix},\text{ and }x_{t}=\begin{bmatrix}x_{u,t}\\ x_{s,t}\end{bmatrix}. (12)
zt+1=Fs​As​Fs−1​zt+Fs​Bs​ut.\displaystyle z_{t+1}=F_{s}A_{s}F_{s}^{-1}z_{t}+F_{s}B_{s}u_{t}. (13)
Original figure
Figure 2: SIG vs. EP-IDM across two examples. Each row compares a representation trained with SIGReg (SIG) against one trained with the endpoint inverse-dynamics (action-reconstruction) loss (EP-IDM). Top row: synthetic open-loop unstable LDS. Bottom row: linearized CartPole. Left: true unstable (vuv_{u}, green) and stable (vsv_{s}, grey) eigenvectors versus the learned predictor’s dominant eigenvector, for SIG (blue) and EP-IDM (dashed orange). Center: closed-loop trajectories under the latent LQR controller from four initial states (circles) for the model trained with SIGReg. Right: closed-loop trajectories under the latent LQR controller for the model trained with the inverse dynamics loss (EP-IDM). For the bottom row x3x_{3}, and x4x_{4} corresponds to the pole angle and angular velocity, respectively. The latent dimension is one for the example in the top row and three for the example in the bottom row.

.

5 Preserving Unstable Directions Through Inverse Dynamics

The added endpoint objective asks the initial and final latent pair to explain which action sequence produced the transition. This encourages the encoder to retain action-induced changes rather than merely predictable passive variation.

5.1 Endpoint Inverse Dynamics

An inverse model reconstructs the stacked action block from endpoint latents. It is trained jointly with the encoder and predictor. In practice this is a soft objective; zero reconstruction loss is not assumed to be attainable for long arbitrary action sequences.

Equation 14
ℒEP​-​IDM​(θ)=𝔼τ∼𝒟​[1H​∑t=0H−1∥[Dθ​(ztτ,ztτ+H)]t−atτ+t∥22],\mathcal{L}_{\mathrm{EP}\text{-}\mathrm{IDM}}(\theta)=\mathbb{E}_{\tau\sim\mathcal{D}}\left[\frac{1}{H}\sum_{t=0}^{H-1}\big\lVert\big[D_{\theta}(z_{t_{\tau}},\,z_{t_{\tau}+H})\big]_{t}-a_{t_{\tau}+t}\big\rVert_{2}^{2}\right], (14)

5.2 Preservation of Reachable Directions

For affine encoding and inverse dynamics, exact endpoint action recovery under rich conditional action excitation forces preservation of the finite-horizon reachable subspace. The condition requires enough endpoint information and rank: arbitrary action blocks cannot be recovered when their dimension exceeds the state or latent capacity. The empirical loss remains useful as a bias even outside this ideal regime.

Equations 15–16
Dθ​(z0,zH)=WH​[z0​zH]+bH​ with ​WH∈ℝm​H×2​d​ and ​bH∈ℝm​H.D_{\theta}(z_{0},z_{H})=W_{H}[z_{0}\;z_{H}]+b_{H}\text{ with }W_{H}\in\mathbb{R}^{mH\times 2d}\text{ and }b_{H}\in\mathbb{R}^{mH}. (15)
ker⁡(F)∩ℛH={0}.\displaystyle\ker(F)\cap\mathcal{R}_{H}=\{0\}. (16)

5.3 Reachability and Unstable Dynamics

Longer-horizon controllability Gramians increasingly weight controllable unstable directions under the paper’s strengthened growth and non-cancellation assumptions. Their dominant subspaces then approach the unstable subspace. Mere reachability without those conditions is not the full theorem, and extending the horizon does not override the endpoint reconstruction rank constraint.

Equations 17–19
‖sin⁡Θ⁡(Φ^,Φ)‖2→0as ​H→∞.\|\sin\Theta(\hat{\Phi},\Phi)\|_{2}\to 0\qquad\text{as }H\to\infty. (17)
λmin​(Πu,H)=min‖v‖2=1⁡v⊤​Πu,H​v≥cu​α2​H,\displaystyle\lambda_{\min}(\Pi_{u,H})=\min_{\|v\|_{2}=1}v^{\top}\Pi_{u,H}v\geq c_{u}\alpha^{2H}, (18)
v⊤​Πu,H​v≥cu​α2​Hfor every ​‖v‖2=1\displaystyle v^{\top}\Pi_{u,H}v\geq c_{u}\alpha^{2H}\qquad\text{for every }\|v\|_{2}=1 (19)

6 Experiments

Experiments compare one- and multi-step prediction, Gaussian regularization and several action-reconstruction variants. Local linear feedback, gradient planning and sampling-based planning test different uses of the same model.

Refer to caption
Figure 3: CartPole, Walker2D, and PointMaze visual-control tasks used in our evaluation. Each row depicts ten frames from a successful trial (initial state: blue border, final state: green border). CartPole: a continuous-action inverted-pendulum balancing task controlled via LQR. Walker2D: a bipedal robot locomotion task controlled via iCEM (Pinneri et al., 2021). PointMaze: a 2-D point-mass robot navigation task planned via CEM.
Original figure
Figure 4: CartPole state norm ‖xt‖2\|x_{t}\|_{2} under the designed latent LQR controller over 300300 steps.

6.1 Setup

CartPole tests an unstable equilibrium, Walker2D a feedback-maintained walking cycle, and PointMaze a stable goal-reaching system. Controllers include latent LQR and nonlinear planning. Data, proprioception and action contexts are specified because they materially affect which control information is available.

Original figure
Figure 5: Walker2D. Phase portraits of right-hip angle vs. angular velocity for GT dynamics, 1SP+EP-IDM, 1SP+SIG, and MSP+SIG. Each trajectory is a 500500-step closed-loop rollout: GT uses real MuJoCo physics, model panels roll out in latent space under an MLP-decoded state, with a SAC (Huang et al., 2022; Haarnoja et al., 2018) policy acting on the decoded observation.

6.2 CartPole: Stabilization Around an Unstable Equilibrium

Endpoint action reconstruction changes reported latent-LQR CartPole success from zero to 100% in the small evaluated setting. Some SIGReg models still succeed with nonlinear replanning, so failure of a local linear controller should not be misreported as failure of every possible planner.

Table 1: CartPole stabilization success rate (%).
Model LQR (SR) LQR (MFS) CEM (SR) GBP (SR)
Ground truth (GT) 100% 1.000 100% 100%
1SP + SIG 0% 0.202 100% 90%
MSP + SIG 0% 0.247 100% 100%
1SP + EP-IDM 100% 0.999 100% 100%
MSP + EP-IDM + SIG 100% 0.996 100% 100%

6.3 Walker2D: Preserving a Walking Gait

Endpoint-reconstruction models better preserve the studied gait’s cyclic structure and support stronger forward locomotion than the compared prediction-only variants. Phase portraits, decoded rollouts and control metrics test complementary aspects of the learned dynamics; the linear equilibrium theorem does not prove this nonlinear orbit result.

Refer to caption
Figure 6: Trajectory frames for Walker2D. Each row shows ten frames uniformly sampled from a 500500-step closed-loop rollout under the SAC policy (Haarnoja et al., 2018), decoded from the trained model’s latent state via an MLP decoder.
Table 2: Walker2D locomotion under latent iCEM (Pinneri et al., 2021). All metrics are averaged over ten trials. Final height is the torso height (m) at the end of the episode.
Model Avg Velocity (m/s) Displacement (m) Final Height (m)
Ground truth (GT) 3.61 14.43 1.21
1SP + SIG 0.08 0.06 0.98
MSP + SIG 0.46 0.66 0.85
1SP + EP-IDM 2.91 9.15 0.92
Refer to caption
Figure 7: Planning trajectory frames for Walker2D. Each row shows ten frames uniformly sampled from a 500500-step closed-loop rollout under latent iCEM planning for the trained models. We select the best trial, with respect to the forward displacement, across the ten trials for the trained models.

6.4 Diagnostics

Vector fields, local rollout error, attraction regions and Lyapunov differences explain why similar-looking goal costs can support different feedback controllers. PointMaze shows that action reconstruction can also preserve useful state geometry in a system without unstable modes. Diagnostics remain empirical over sampled states.

Refer to caption
Figure 8: CartPole diagnostics. Left: Phase portraits of ground-truth and learned vector fields. Center: H=3H=3 open-loop prediction error across (θ,θ˙)(\theta,\dot{\theta}). Right: Zero-action planning cost.
Table 3: PointMaze navigation success rate (%).
Encoder Objective CEM (SR) LQR (SR) GBP (SR)
GT 100% 80% 90.0%*
From scratch 1SP + IDM 100% 70% 100%
MSP + MS-IDM 100% 80% 90.0%
MSP + EP-IDM 90.0% 80% 90.0%
1SP + EP-IDM 100% 80% 80.0%
MSP + SIG 90.0% 60% 100%
1SP + IDM + MS-IDM 90.0% 80% 90.0%
DINOv2 MSP 50.0% 20% 80%
iBOT MSP 100% 80% 80%
DINO-WM 80% 80% 60%
  • *

    Because MuJoCo is non-differentiable, we estimate gradients with SPSA using two rollouts per optimization step (Spall, 1998).

Refer to caption
Figure 9: Local stability on CartPole. First two panels: Empirical region of attraction of the learned LQR controller when deployed in the ground-truth system. The dashed line indicates the ground-truth LQR boundary. Last two panels: Lyapunov decrease Δ​V=V⁡(x′)−V⁡(x)\Delta V=V(x^{\prime})-V(x) under the latent LQR controller.

7 Conclusion and Future Work

The central separation is between noncollapsed predictability and information sufficient for feedback stabilization. Endpoint inverse dynamics supplies a control-aware learning signal, with an exact theorem only under restrictive linear and rank conditions. Broader nonlinear stability guarantees and combinations with planning-geometry objectives remain open.

8 Acknowledgments

Acknowledgments and funding disclosures are preserved as source context. They do not change the theorem assumptions or experimental scope.

Appendix A Appendix

The appendix retains additional rollouts, frozen-encoder tests, implementation details, ablations, linear counterexample constructions and complete proofs.

A.1 Additional Experimental Validation (Extending Section 6 )

Additional diagnostics distinguish rollout fidelity, informative latent costs and stable local feedback. None is a substitute for checking the physical closed-loop trajectory.

A.1.1 CartPole

CartPole sequences and attraction-region plots compare models trained from scratch and models with frozen visual features. Endpoint reconstruction often helps, while a visually plausible goal cost alone may still yield an unstable latent-LQR controller.

Refer to caption
Figure 10: CartPole trajectory visualization for the trained models under the latent LQR controller.
Refer to caption
Figure 11: MSP+EP-IDM+SIG model on CartPole. Left: Phase portrait of ground-truth vs. learned vector fields. Center: H=3H=3 open-loop prediction error. Right: Zero-action planning cost.
Refer to caption
Figure 12: Local stability of MSP+EP-IDM+SIG on CartPole. Left: Empirical region of attraction (ROA) of the learned LQR policy in the GT environment. Dashed line: GT LQR boundary. Right: Lyapunov decrease Δ​V=V⁡(x′)−V⁡(x)\Delta V=V(x^{\prime})-V(x) under the learned policy using the GT DARE solution.
Refer to caption
Figure 13: MSP+IBOT+PR-EP-IDM model on CartPole. Panel descriptions are in Fig.  11.
Refer to caption
Figure 14: Local stability of MSP+IBOT+PR-EP-IDM on CartPole. Panel descriptions are the same as those in Fig.  12.
Refer to caption
Figure 15: MSP+IBOT+PR-SIG model on CartPole. Panel descriptions are the same as those in Fig.  11.
Refer to caption
Figure 16: MSP+IBOT model on CartPole. Panel descriptions are the same as those in Fig.  11.
Refer to caption
Figure 17: Local stability of MSP+IBOT on CartPole. Left: Empirical region of attraction (ROA) of the learned LQR policy in the GT environment. Panel descriptions are as those in Fig.  12.
Refer to caption
Figure 18: DINO-WM model on CartPole. Panel descriptions are the same as those in Fig.  11.
Table 4: CartPole CEM diagnostic metrics (action context 1). 1/L21/L^{2}: inverse squared mean stabilization length. FE: final error ‖xT‖\|x_{T}\|. PR denotes a learned projector on top of the frozen encoder.
Encoder Objective CEM (1/L21/L^{2}) CEM (FE)
GT 0.0003 0.30202
From scratch 1SP + SIG 0.0003 0.19956
1SP + IDM + MS-IDM 0.0003 0.43608
1SP + EP-IDM 0.0003 0.13683
MSP + SIG 0.0003 0.07667
MSP + EP-IDM + SIG 0.0003 0.11985
DINOv2 1SP 0.0038 4.60613
MSP 0.0031 5.09216
MSP + PR-SIG 0.0003 0.19056
MSP + PR-IDM 0.0331 4.68120
MSP + PR-EP-IDM 0.0004 1.57507
iBOT 1SP 0.0003 0.09306
1SP + PR-EP-IDM 0.0004 0.47275
MSP 0.0073 3.49471
MSP + PR-SIG 0.0004 1.42928
MSP + PR-IDM 0.0567 5.11192
MSP + PR-EP-IDM 0.0003 0.62596
DINO-WM 0.0005 0.9777

A.1.2 PointMaze

PointMaze does not require preserving an unstable subspace. Its goal-distance maps show whether the representation gives the nonlinear planner an informative route cost, providing a complementary test of action-relevant geometry.

Refer to caption
Figure 19: PointMaze visualization for our trained models under CEM planning.
Refer to caption
Figure 20: Latent goal-distance fields ‖z−zg‖\|z-z_{\mathrm{g}}\| across the PointMaze U-maze for four trained world models.

A.1.3 Closed-Loop Prediction Error

Long closed-loop prediction traces show that multi-step training alone need not prevent drift. Combining action reconstruction and marginal regularization can improve different aspects of the model; one-step error is not a complete summary of either.

Original figure
Figure 21: kk-step closed-loop prediction error along trajectories generated by the ground-truth LQR controller. Curves show the physical-state error between the ground-truth trajectory and the trajectory decoded from each latent rollout.

A.2 Pre-trained encoders

A learned projector can retain useful control information present in a frozen backbone, but cannot reconstruct information the backbone has already removed. The iBOT and DINOv2 comparisons illustrate this limitation and the distinction between LQR and nonlinear-planning success.

Table 5: CartPole control performance with frozen pretrained encoders. SR: success rate. MFS: mean fraction of steps within the success threshold. PR denotes a learned projector. DINO-WM (Zhou et al., 2024) is an external baseline.
Encoder Objective LQR (SR) LQR (MFS) CEM (SR) GBP (SR)
DINOv2 1SP 0% 0.301 0% 0%
MSP 0% 0.232 0% 0%
1SP + PR-EP-IDM 0% 0.025 20% 0%
MSP + PR-SIG 0% 0.460 100% 0%
MSP + PR-IDM 0% 0.455 0% 0%
MSP + PR-EP-IDM 0% 0.513 40% 10%
iBOT 1SP 0% 0.249 100% 80%
MSP 0% 0.228 0% 0%
1SP + PR-EP-IDM 0% 0.263 100% 80%
MSP + PR-SIG 50% 0.571 80% 80%
MSP + PR-IDM 0% 0.405 0% 0%
MSP + PR-EP-IDM 100% 0.998 100% 0%
DINO-WM 0% 0.019 40% 10%
Original figure
Figure 22: CartPole state norm ‖xt‖2\|x_{t}\|_{2} under the learned latent LQR controller over 300300 steps. Left: DINOv2. Right: iBOT.

A.3 Additional Details on the Implementation of our Experiments

The reproducibility section specifies perception, action decoding, datasets and controller budgets. These choices constrain the meaning of the nonlinear evidence.

A.3.1 Model Architecture

The from-scratch encoder combines frame-difference information and proprioception into a compact latent. Predictor and action-decoder architectures are retained. Calling the task visual does not mean every experiment excludes measured physical-state information.

Table 6: Task-specific observation and action dimensions. The frame skip is the number of simulator steps represented by one model transition.
Task Image size Proprioception Action Frame skip Training trajectories
CartPole 128×128128\times 128 44 11 55 787787
Walker2D 64×6464\times 64 1717 66 55 3,2003,200
PointMaze 64×6464\times 64 44 22 55 1,6001,600
Table 7: World-model training hyperparameters.
Hyperparameter Value Hyperparameter Value
Optimizer AdamW Epochs 200200
Batch size 6464 Prediction horizon HH 33
Initial learning rate 10−410^{-4} Minimum learning rate 10−610^{-6}
Weight decay 10−310^{-3} Warm-up epochs 55
Gradient-norm threshold 11 Latent dimension 192192
Random seed 4242 Active loss weights 11

A.3.2 Task and Dataset Details

Dataset mixtures include passive, random and stabilizing behavior, with separate training, validation and test trajectories. The small number of tested initial conditions limits interpreting perfect success as a broad probabilistic guarantee.

A.3.3 Planning and Control Details

LQR relies on local linearization; CEM, improved CEM and gradient planning optimize finite action sequences. The Walker2D improved-CEM setup uses the stated policy warm start, which is an important additional resource.

Table 8: CEM and improved-CEM planning hyperparameters. Here, “execute” denotes the number of actions applied before replanning.
Task Planner Horizon Execute Population Elites Iterations
CartPole CEM 1010 11 300300 3030 1010
Walker2D iCEM 33 11 200200 2020 55
PointMaze CEM 2525 2525 300300 3030 1010

A.3.4 Sketched Isotropic Gaussian Regularization (SIGReg)

SIGReg matches sampled scalar projections of the latent distribution to a Gaussian reference. This discourages global collapse but does not directly encode a detectability or controllability constraint.

Equations
LSIG​(θ)=1K​∑ℓ=1K(∑j=1nωg⁡(ωj)​[(Re⁡φ^ℓ​(ωj)−φ𝒩​(ωj))2+(Im⁡φ^ℓ​(ωj))2]∑j=1nωg⁡(ωj)).\displaystyle L_{\mathrm{SIG}}(\theta)=\frac{1}{K}\sum_{\ell=1}^{K}\left(\frac{\sum_{j=1}^{n_{\omega}}g(\omega_{j})\Big[\left(\operatorname{Re}\hat{\varphi}_{\ell}(\omega_{j})-\varphi_{\mathcal{N}}(\omega_{j})\right)^{2}+\left(\operatorname{Im}\hat{\varphi}_{\ell}(\omega_{j})\right)^{2}\Big]}{\sum_{j=1}^{n_{\omega}}g(\omega_{j})}\right).
φ^ℓ​(ω)\displaystyle\hat{\varphi}_{\ell}(\omega) =1M∑i=1Mei​ω​(wℓ⊤​zi),φ𝒩(ω)=e−ω2/2=:g(ω),wℓ∼i.i.d.Unif(𝕊d−1).\displaystyle=\frac{1}{M}\sum_{i=1}^{M}e^{\mathrm{i}\omega(w_{\ell}^{\top}z_{i})},\qquad\varphi_{\mathcal{N}}(\omega)=e^{-\omega^{2}/2}=:g(\omega),\qquad w_{\ell}\overset{\text{i.i.d.}}{\sim}\mathrm{Unif}\bigl(\mathbb{S}^{d-1}\bigr).

A.3.5 Additional Action Reconstruction Losses

One-step, multi-step-trajectory and endpoint inverse dynamics expose the action decoder to different latent information. Their losses are not equivalent: an entire intermediate trajectory may reveal actions that endpoints alone cannot distinguish.

Equations 20–24
DMS​-​IDM,θ​(ztτ,…,ztτ+H)\displaystyle D_{\mathrm{MS}\text{-}\mathrm{IDM},\theta}(z_{t_{\tau}},\ldots,z_{t_{\tau}+H}) =𝐚^τ,H∈ℝm​H,\displaystyle=\hat{\mathbf{a}}_{\tau,H}\in\mathbb{R}^{mH}, (20)
ℒMS​-​IDM​(θ)=𝔼τ∼𝒟​[1H​∑t=0H−1‖[DMS​-​IDM,θ​(ztτ,…,ztτ+H)]t−atτ+t‖22].\displaystyle\mathcal{L}_{\mathrm{MS}\text{-}\mathrm{IDM}}(\theta)=\mathbb{E}_{\tau\sim\mathcal{D}}\left[\frac{1}{H}\sum_{t=0}^{H-1}\left\|\left[D_{\mathrm{MS}\text{-}\mathrm{IDM},\theta}\bigl(z_{t_{\tau}},\ldots,z_{t_{\tau}+H}\bigr)\right]_{t}-a_{t_{\tau}+t}\right\|_{2}^{2}\right]. (21)
DIDM,θ​(ztτ+t,ztτ+t+1)\displaystyle D_{\mathrm{IDM},\theta}(z_{t_{\tau}+t},z_{t_{\tau}+t+1}) =a^tτ+t∈ℝm.\displaystyle=\hat{a}_{t_{\tau}+t}\in\mathbb{R}^{m}. (22)
ℒIDM​(θ)=𝔼τ∼𝒟​[1H​∑t=0H−1‖DIDM,θ​(ztτ+t,ztτ+t+1)−atτ+t‖22].\displaystyle\mathcal{L}_{\mathrm{IDM}}(\theta)=\mathbb{E}_{\tau\sim\mathcal{D}}\left[\frac{1}{H}\sum_{t=0}^{H-1}\left\|D_{\mathrm{IDM},\theta}\bigl(z_{t_{\tau}+t},z_{t_{\tau}+t+1}\bigr)-a_{t_{\tau}+t}\right\|_{2}^{2}\right]. (23)
ℒ¯IDM​(θ)=λ1​ℒEP​-​IDM​(θ)+λ2​ℒMS​-​IDM​(θ)+λ3​ℒIDM​(θ).\displaystyle\bar{\mathcal{L}}_{\mathrm{IDM}}(\theta)=\lambda_{1}\mathcal{L}_{\mathrm{EP}\text{-}\mathrm{IDM}}(\theta)+\lambda_{2}\mathcal{L}_{\mathrm{MS}\text{-}\mathrm{IDM}}(\theta)+\lambda_{3}\mathcal{L}_{\mathrm{IDM}}(\theta). (24)

A.4 Ablations

The ablations vary inverse-dynamics formulation, action-history length, proprioceptive inputs and regularization. They qualify attributing every gain to the endpoint architecture and demonstrate interactions with available observation information.

Table 9: CartPole ablation over inverse-dynamics formulations for an encoder trained from scratch.
Objective LQR (SR) LQR (MFS) CEM (SR) GBP (SR)
1SP + IDM + MS-IDM 100% 0.999 100% 100%
1SP + EP-IDM 100% 0.999 100% 100%
MSP + EP-IDM + SIG 100% 0.996 100% 100%
Table 10: CartPole success rates (%) with action contexts one and five. PR denotes a learned projector on top of the frozen encoder.
LQR (SR) CEM (SR) GBP (SR)
Encoder Objective Context 1 Context 5 Context 1 Context 5 Context 1 Context 5
From scratch 1SP + SIG 0 0 100 100 90 90
1SP + IDM + MS-IDM 100 100 100 100 100 100
1SP + EP-IDM 100 100 100 100 100 100
MSP + SIG 0 40 100 100 100 100
DINOv2 MSP 0 0 0 0 0 0
DINOv2 MSP + PR-SIG 0 0 100 60 0 0
iBOT MSP 0 0 0 100 0 60
MSP + PR-SIG 50 40 80 90 80 60
Table 11: PointMaze success rates (%) with action contexts one and five.
LQR (SR) CEM (SR) GBP (SR)
Encoder Objective Context 1 Context 5 Context 1 Context 5 Context 1 Context 5
From scratch 1SP + IDM 70 80 100 90 100 80
1SP + EP-IDM 80 80 100 100 80 80
1SP + IDM + MS-IDM 80 70 90 100 90 90
MSP + SIG 60 80 90 70 100 60
MSP + MS-IDM 80 80 100 100 90 90
MSP + EP-IDM 80 80 90 90 90 100
DINOv2 MSP 20 10 50 60 80 50
iBOT MSP 80 80 100 80 80 60
Table 12: CartPole ablation: Action context one, only proprioceptive information (state).
Predictor Objective LQR (SR) LQR (MFS) CEM (SR) CEM (held) GBP (SR/held)
MLP MSP + SIG 20% 0.571 20% 20% 0% / 0%
1SP + EP-IDM 0% 0.204 100% 100% 0% / 0%
Transformer MSP + SIG 20% 0.461 0% 0% 0% / 0%
1SP + EP-IDM 80% 0.869 100% 100% 100% / 100%
Table 13: CartPole ablation with a frozen iBOT encoder and action context five, with and without proprioceptive inputs. PR denotes the learned projector. SR: success rate, held: fraction of trials in which stabilization is preserved.
Objective Proprioception LQR (SR) CEM (SR) CEM (held) GBP (SR) GBP (held)
MSP Yes 0% 100% 100% 60% 40%
No 0% 0% 0% 0% 0%
MSP + PR-SIG Yes 40% 90% 90% 60% 60%
No 0% 0% 0% 0% 0%
MSP + PR-IDM Yes 100% 80% 80% 70% 70%
No 0% 0% 0% 0% 0%
Table 14: Ablation on the encoder gain ‖Δ​zt‖/‖Δ​xt‖\|\Delta z_{t}\|/\|\Delta x_{t}\| as a function of initial-state perturbation magnitude ‖x0‖\|x_{0}\|. We start from a perturbed upright equilibrium, the cart pole is stepped with zero action for 60 steps. We measure the mean latent displacement ‖Δ​zt‖=‖zt−zt−1‖\|\Delta z_{t}\|=\|z_{t}-z_{t-1}\| divided by the mean physical-state displacement ‖Δ​xt‖=‖xt−xt−1‖\|\Delta x_{t}\|=\|x_{t}-x_{t-1}\| .
Encoder ‖x0‖=0.005\|x_{0}\|{=}0.005 ‖x0‖=0.01\|x_{0}\|{=}0.01 ‖x0‖=0.02\|x_{0}\|{=}0.02 ‖x0‖=0.05\|x_{0}\|{=}0.05 ‖x0‖=0.10\|x_{0}\|{=}0.10 ‖x0‖=0.20\|x_{0}\|{=}0.20
iBOT 1.07×\times 1.11×\times 1.07×\times 1.07×\times 1.02×\times 1.03×\times
DINOv2 25.45×\times 26.07×\times 26.38×\times 25.67×\times 21.83×\times 25.12×\times
Table 15: Fixed-point residual of the learned predictor at the encoded upright CartPole equilibrium under zero action. The residual measures the one-step drift from this encoded equilibrium. PR denotes a learned projector on top of the frozen encoder.
Encoder Objective ‖P⁡(zeq,0)−zeq‖\|P(z_{\mathrm{eq}},0)-z_{\mathrm{eq}}\|
DINOv2 1SP 5.569
MSP 7.215
1SP + PR-EP-IDM 0.4026
MSP + PR-SIG 2.515
MSP + PR-EP-IDM 0.343
iBOT 1SP 0.059
MSP 0.101
1SP + PR-EP-IDM 0.091
MSP + PR-SIG 0.641
MSP + PR-EP-IDM 0.129

A.5 Additional Details on Examples of Section 4

Explicit linear systems and perceptual mixing maps show how an encoder can discard unstable coordinates without failing its training objectives. A stable-system example separates control relevance from instability specifically. These are constructed counterexamples, not empirical estimates of how often collapse occurs.

Equations 25–31
A¯=diag⁡(1.25,0.85)​ and ​B¯=[10.6].\displaystyle\bar{A}=\operatorname{diag}(1.25,0.85)\text{ and }\bar{B}=\begin{bmatrix}1\\ 0.6\end{bmatrix}. (25)
Tϑ=[cos⁡ϑ−sin⁡ϑsin⁡ϑcos⁡ϑ],ϑ=0.6,A=Tϑ​A¯​Tϑ−1, and ​B=Tϑ​B¯.\displaystyle T_{\vartheta}=\begin{bmatrix}\cos\vartheta&-\sin\vartheta\\ \sin\vartheta&\cos\vartheta\end{bmatrix},\;\vartheta=0.6,\;A=T_{\vartheta}\bar{A}T_{\vartheta}^{-1},\text{ and }B=T_{\vartheta}\bar{B}. (26)
A=[1.1220.1860.1860.978]∈ℝ2×2​ and ​B=[0.4871.060]∈ℝ2×1.\displaystyle A=\begin{bmatrix}1.122&0.186\\ 0.186&0.978\end{bmatrix}\in\mathbb{R}^{2\times 2}\text{ and }B=\begin{bmatrix}0.487\\ 1.060\end{bmatrix}\in\mathbb{R}^{2\times 1}. (27)
C=[0.069−0.0810.3530.064−0.2950.2220.7180.581−0.388−0.776−0.3430.025]∈ℝ6×2.\displaystyle C=\begin{bmatrix}0.069&-0.081\\ 0.353&0.064\\ -0.295&0.222\\ 0.718&0.581\\ -0.388&-0.776\\ -0.343&0.025\end{bmatrix}\in\mathbb{R}^{6\times 2}. (28)
λu=1.25​(unstable),vu=[0.8250.565]​ and ​λs=0.85​(stable),vs=[−0.5650.825].\displaystyle\lambda_{u}=1.25\ (\textbf{unstable}),\ v_{u}=\begin{bmatrix}0.825\\ 0.565\end{bmatrix}\text{ and }\lambda_{s}=0.85\ (\textbf{stable}),\ v_{s}=\begin{bmatrix}-0.565\\ 0.825\end{bmatrix}. (29)
λs​1=0.25​(stable 1),vs​1=[0.8250.565]​ and ​λs​2=0.85​(stable 2),vs​2=[−0.5650.825].\displaystyle\lambda_{s1}=0.25\ (\textbf{stable 1}),\ v_{s1}=\begin{bmatrix}0.825\\ 0.565\end{bmatrix}\text{ and }\lambda_{s2}=0.85\ (\textbf{stable 2}),\ v_{s2}=\begin{bmatrix}-0.565\\ 0.825\end{bmatrix}. (30)
A¯=diag⁡(0.25,0.85)​ and ​A=Tϑ​A¯​Tϑ−1=[0.441−0.280−0.2800.659],\displaystyle\bar{A}=\operatorname{diag}(0.25,0.85)\text{ and }A=T_{\vartheta}\bar{A}T_{\vartheta}^{-1}=\begin{bmatrix}0.441&-0.280\\ -0.280&0.659\end{bmatrix}, (31)
Original figure
Figure 23: Synthetic open-loop stable LDS. This example considers a two-dimensional open-loop stable system. Panels descriptions are the same as those in Fig.  2.
Equations 32–35
x˙=Ac​x+Bc​a,\displaystyle\dot{x}=A_{c}x+B_{c}a, (32)
Ac\displaystyle A_{c} =[010000−m​gM0000100(M+m)​gM​ℓ0]​ and ​Bc=[01M0−1M​ℓ].\displaystyle=\begin{bmatrix}0&1&0&0\\ 0&0&-\frac{mg}{M}&0\\ 0&0&0&1\\ 0&0&\frac{(M+m)g}{M\ell}&0\end{bmatrix}\text{ and }B_{c}=\begin{bmatrix}0\\ \frac{1}{M}\\ 0\\ -\frac{1}{M\ell}\end{bmatrix}. (33)
A=[1.0000.020−0.0002−0.00000.0001.000−0.0196−0.00020.0000.0001.00430.02000.0000.0000.43231.0043]∈ℝ4×4​ and ​B=[0.00020.0200−0.0004−0.0401]∈ℝ4×1.\displaystyle A=\begin{bmatrix}1.000&0.020&-0.0002&-0.0000\\ 0.000&1.000&-0.0196&-0.0002\\ 0.000&0.000&1.0043&0.0200\\ 0.000&0.000&0.4323&1.0043\end{bmatrix}\in\mathbb{R}^{4\times 4}\text{ and }B=\begin{bmatrix}0.0002\\ 0.0200\\ -0.0004\\ -0.0401\end{bmatrix}\in\mathbb{R}^{4\times 1}. (34)
C=[0.046−0.0680.2560.056−0.1950.1850.5220.503−0.256−0.647−0.2500.022−0.847−0.112−0.499−0.389−0.198−0.1620.1650.553−0.0470.699−0.2660.1870.3290.048−0.298−0.489−0.1670.113−0.404−0.111]∈ℝ8×4.\displaystyle C=\begin{bmatrix}0.046&-0.068&0.256&0.056\\ -0.195&0.185&0.522&0.503\\ -0.256&-0.647&-0.250&0.022\\ -0.847&-0.112&-0.499&-0.389\\ -0.198&-0.162&0.165&0.553\\ -0.047&0.699&-0.266&0.187\\ 0.329&0.048&-0.298&-0.489\\ -0.167&0.113&-0.404&-0.111\end{bmatrix}\in\mathbb{R}^{8\times 4}. (35)
Refer to caption
Figure 24: Sign of Δ​V=V⁡(x′)−V⁡(x)\Delta V=V(x^{\prime})-V(x) after one closed-loop step under the learned latent LQR controller, where V⁡(x)=x⊤​S​xV(x)=x^{\top}Sx and S≻0S\succ 0 is the ground-truth Lyapunov function matrix. We also depict the Δ​V=0\Delta V=0 contour (separatrix), i.e., the boundary separating states where the learned controller decreases VV (blue) from those where it does not (red).

A.6 Proofs of Our Theoretical Guarantees

The proofs retain the supporting rank, subspace and perturbation results. Their linear-algebra conditions should accompany any statement of exact preservation or convergence.

A.6.1 Supporting Lemmas

Exact affine reconstruction on an open set forces an observation map to have a left inverse on the reconstructed coordinates. Matrix perturbation results subsequently relate dominant Gramian eigenspaces to the unstable subspace.

Equations 36–37
(L​T−I)​u+(L​c+b)=0\displaystyle(LT-I)u+(Lc+b)=0 (36)
‖sin⁡Θ⁡(Φ^,Φ)‖2≤2​‖E‖2δ.\displaystyle\|\sin\Theta(\hat{\Phi},\Phi)\|_{2}\leq\frac{2\|E\|_{2}}{\delta}. (37)

A.6.2 Proof of Lemma 1

If an invisible unstable direction separates two states, a controller receiving identical encoded histories cannot distinguish them, while their difference grows. Detectability rules out this obstruction and supplies the converse linear feedback argument under the stated system assumptions.

Equations 38–40
F​At​v=λt​F​v=0.\displaystyle FA^{t}v=\lambda^{t}Fv=0. (38)
x^t+1\displaystyle\hat{x}_{t+1} =A​x^t+B​at+L⁡(z~t−F​x^t),\displaystyle=A\hat{x}_{t}+Ba_{t}+L(\tilde{z}_{t}-F\hat{x}_{t}),
at\displaystyle a_{t} =K​x^t.\displaystyle=K\hat{x}_{t}. (39)
xt+1=(A+B​K)​xt−B​K​et.\displaystyle x_{t+1}=(A+BK)x_{t}-BKe_{t}. (40)

A.6.3 Proof of Theorem 1

Unrolling linear dynamics expresses endpoint changes through the finite-horizon controllability matrix. Exact recovery of a richly excited action block forces its encoded image to retain rank, which prevents the encoder from annihilating reachable directions. The required rank may be impossible at large action-block dimension.

Equations 41–43
xt+H=AH​xt+𝒞H​𝐚t\displaystyle x_{t+H}=A^{H}x_{t}+\mathcal{C}_{H}\mathbf{a}_{t} (41)
zt+H=F​AH​xt+F​𝒞H​𝐚t,H+b.\displaystyle z_{t+H}=FA^{H}x_{t}+F\mathcal{C}_{H}\mathbf{a}_{t,H}+b. (42)
F​𝒞H​q=F​v=0.\displaystyle F\mathcal{C}_{H}q=Fv=0. (43)

A.6.4 Proof of Theorem 2

The Gramian proof separates stable and unstable contributions and bounds their growth and coupling before applying subspace perturbation arguments. Uniform expansion, sufficient excitation and non-cancellation conditions are material assumptions; they are stronger than merely stating that an unstable mode is controllable.

Equations 44–61
A=[AuΔ0As],B=[BuBs].A=\begin{bmatrix}A_{u}&\Delta\\ 0&A_{s}\end{bmatrix},\qquad B=\begin{bmatrix}B_{u}\\ B_{s}\end{bmatrix}. (44)
Aj​B=[ΓjAsj​Bs].A^{j}B=\begin{bmatrix}\Gamma_{j}\\ A_{s}^{j}B_{s}\end{bmatrix}. (45)
σmin​(Au)\displaystyle\sigma_{\min}(A_{u}) ≥α,‖Asj‖2≤C1​βj,‖Auj‖2≤C2​γj,‖Γj‖2≤C3​γj,∀j≥0, and ​γ​β<α2.\displaystyle\geq\alpha,\;\;\|A_{s}^{j}\|_{2}\leq C_{1}\beta^{j},\;\;\|A_{u}^{j}\|_{2}\leq C_{2}\gamma^{j},\;\;\|\Gamma_{j}\|_{2}\leq C_{3}\gamma^{j},\forall j\geq 0,\;\text{ and }\;\gamma\beta<\alpha^{2}. (46)
λmin​(Πu,H)≥cu​α2​H, for ​H≥H0.\displaystyle\lambda_{\min}(\Pi_{u,H})\geq c_{u}\alpha^{2H},\text{ for }H\geq H_{0}. (47)
λmin​(Πu,H)\displaystyle\lambda_{\min}(\Pi_{u,H}) ≥cu​α2​H,‖Πs,H‖2≤cs, and ​‖Πu​s,H‖2≤cu​s​∑j=0H−1(γ​β)j,\displaystyle\geq c_{u}\alpha^{2H},\;\|\Pi_{s,H}\|_{2}\leq c_{s},\text{ and }\|\Pi_{us,H}\|_{2}\leq c_{us}\sum_{j=0}^{H-1}(\gamma\beta)^{j}, (48)
λmin​(Πu,H)\displaystyle\lambda_{\min}(\Pi_{u,H}) ≥cu​α2​H,\displaystyle\geq c_{u}\alpha^{2H},
‖Πs,H‖2\displaystyle\|\Pi_{s,H}\|_{2} ≤cs,\displaystyle\leq c_{s},
‖Πu​s,H‖2\displaystyle\|\Pi_{us,H}\|_{2} ≤cu​s​∑j=0H−1(γ​β)j.\displaystyle\leq c_{us}\sum_{j=0}^{H-1}(\gamma\beta)^{j}. (49)
‖Πs,H‖2\displaystyle\|\Pi_{s,H}\|_{2} ≤∑j=0H−1‖Asj​Bs‖22≤C12​‖Bs‖22​∑j=0∞β2​j=C12​‖Bs‖221−β2,\displaystyle\leq\sum_{j=0}^{H-1}\|A_{s}^{j}B_{s}\|_{2}^{2}\leq C_{1}^{2}\|B_{s}\|_{2}^{2}\sum_{j=0}^{\infty}\beta^{2j}=\frac{C_{1}^{2}\|B_{s}\|_{2}^{2}}{1-\beta^{2}},
‖Πu​s,H‖2\displaystyle\|\Pi_{us,H}\|_{2} ≤∑j=0H−1‖Γj‖2​‖Asj​Bs‖2≤C3​C1​‖Bs‖2​∑j=0H−1(γ​β)j.\displaystyle\leq\sum_{j=0}^{H-1}\|\Gamma_{j}\|_{2}\|A_{s}^{j}B_{s}\|_{2}\leq C_{3}C_{1}\|B_{s}\|_{2}\sum_{j=0}^{H-1}(\gamma\beta)^{j}. (50)
σi(Ξ⊤Ξ^)=cos(θi), for all i=1,…,r,\displaystyle\sigma_{i}(\Xi^{\top}\hat{\Xi})=\cos(\theta_{i}),\text{ for all }i=1,\dots,r, (51)
sin⁡Θ⁡(Φ^,Φ):=diag⁡(sin⁡(θ1),…,sin⁡(θr)).\displaystyle\sin\Theta(\hat{\Phi},\Phi):=\operatorname{diag}\bigl(\sin(\theta_{1}),\dots,\sin(\theta_{r})\bigr). (52)
‖sin⁡Θ⁡(Φ^,Φ)‖2:=maxi=1,…,r⁡sin⁡(θi)=‖PΦ^−PΦ‖2,\displaystyle\bigl\|\sin\Theta(\hat{\Phi},\Phi)\bigr\|_{2}:=\max_{i=1,\dots,r}\sin(\theta_{i})=\bigl\|P_{\hat{\Phi}}-P_{\Phi}\bigr\|_{2}, (53)
ΠH=[Πu,H000]⏟M1+[0Πu​s,HΠu​s,H⊤Πs,H]⏟M2.\displaystyle\Pi_{H}=\underbrace{\begin{bmatrix}\Pi_{u,H}&0\\ 0&0\end{bmatrix}}_{M_{1}}+\underbrace{\begin{bmatrix}0&\Pi_{us,H}\\ \Pi_{us,H}^{\top}&\Pi_{s,H}\end{bmatrix}}_{M_{2}}. (54)
‖M2‖2≤‖Πu​s,H‖2+‖Πs,H‖2≤cu​s​∑j=0H−1(γ​β)j+cs.\displaystyle\|M_{2}\|_{2}\leq\|\Pi_{us,H}\|_{2}+\|\Pi_{s,H}\|_{2}\leq c_{us}\sum_{j=0}^{H-1}(\gamma\beta)^{j}+c_{s}. (55)
‖sin⁡Θ⁡(Φ^,Φ)‖2\displaystyle\|\sin\Theta(\hat{\Phi},\Phi)\|_{2} ≤2​‖M2‖2cu​α2​H≤2​(cu​s​∑j=0H−1(γ​β)j+cs)cu​α2​H.\displaystyle\leq\frac{2\|M_{2}\|_{2}}{c_{u}\alpha^{2H}}\leq\frac{2\left(c_{us}\sum_{j=0}^{H-1}(\gamma\beta)^{j}+c_{s}\right)}{c_{u}\alpha^{2H}}. (56)
‖sin⁡Θ⁡(Φ^,Φ)‖2≤2​(cu​s​∑j=0H−1(γ​β)j+cs)/(cu​α2​H).\bigl\|\sin\Theta(\hat{\Phi},\Phi)\bigr\|_{2}\leq 2\left(c_{us}\displaystyle\sum_{j=0}^{H-1}(\gamma\beta)^{j}+c_{s}\right)/\left(c_{u}\alpha^{2H}\right). (57)
‖sin⁡Θ⁡(Φ^,Φ)‖2≲(qH+1)/α2​H.\bigl\|\sin\Theta(\hat{\Phi},\Phi)\bigr\|_{2}\lesssim(q^{H}+1)/\alpha^{2H}. (58)
‖sin⁡Θ⁡(Φ^,Φ)‖2≲H/α2​H.\bigl\|\sin\Theta(\hat{\Phi},\Phi)\bigr\|_{2}\lesssim H/\alpha^{2H}. (59)
‖sin⁡Θ⁡(Φ^,Φ)‖2≲1/α2​H.\bigl\|\sin\Theta(\hat{\Phi},\Phi)\bigr\|_{2}\lesssim 1/\alpha^{2H}. (60)
‖sin⁡Θ⁡(Φ^,Φ)‖2→0​ as ​H→∞,\bigl\|\sin\Theta(\hat{\Phi},\Phi)\bigr\|_{2}\to 0\text{ as }H\to\infty, (61)
Concept demonstration

Open image ↗