{"id":"arxiv:2605.26379","title":"When Does LeJEPA Learn a World Model?","date":"2026-05-25","section":"Self-supervised learning and vision","quality":0.621282,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/","source_url":"https://arxiv.org/abs/2605.26379","text_available":true,"authors":["David Klindt","Yann LeCun","Randall Balestriero"],"outgoing":[{"source":"arxiv:2605.26379","target":"arxiv:2104.14294","evidence":"title","id":"arxiv:2104.14294","title":"Emerging Properties in Self-Supervised Vision Transformers","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4372186d141ca514b08e93bb3b1e26fdbcba9ce62187fb772c14bf223160c06d/"},{"source":"arxiv:2605.26379","target":"arxiv:2105.04906","evidence":"id","id":"arxiv:2105.04906","title":"VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/822b6414fb02bd6c746e0d559fd3dcc60766476142ebae7795b02bc6c45223ce/"},{"source":"arxiv:2605.26379","target":"arxiv:2301.08243","evidence":"title","id":"arxiv:2301.08243","title":"Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/475bac0a3b29edd42642bbe10d1a41bfb18a3022ac95b9b1a3db9e5d5d0b95fe/"},{"source":"arxiv:2605.26379","target":"arxiv:2511.08544","evidence":"id","id":"arxiv:2511.08544","title":"LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e2c5835a26dabc321e6364380a37c47c10425ee830ec522564b0604733025a94/"},{"source":"arxiv:2605.26379","target":"openreview:BZ5a1r-kVsf","evidence":"url","id":"openreview:BZ5a1r-kVsf","title":"A Path Towards Autonomous Machine Intelligence (LeCun, 2022)","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b7a7789ddc6c70bf2f3e56e6df28529a3eefac2506f6104eb6b2460f5571d657/"}],"incoming":[],"text_parts":[{"start":0,"end":15643,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html"},{"start":15643,"end":31228,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html"},{"start":31228,"end":46970,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-3.html"},{"start":46970,"end":62462,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-4.html"},{"start":62462,"end":78324,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html"},{"start":78324,"end":93753,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html"},{"start":93753,"end":109302,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html"},{"start":109302,"end":124914,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-8.html"},{"start":124914,"end":129109,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-9.html"}],"contents":[{"title":"When Does LeJEPA Learn a World Model?","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"Abstract","offset":41,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"1 Introduction","offset":1349,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"Contributions.","offset":3597,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"2 Related Work","offset":4510,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"Representation Learning.","offset":4529,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"World Models.","offset":5354,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"Identifiability.","offset":6344,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"3 The World and the Learner","offset":7387,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"3.1 The World","offset":8741,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"3.1.1 The Gaussian World","offset":9818,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"3.2 The Learner: LeJEPA","offset":11432,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"4 Spectral Analysis of the World","offset":13346,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"Transition Operator.","offset":13484,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"Gaussian World: Hermite Polynomials.","offset":14405,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"General Case: Sturm-Liouville Theory.","offset":15598,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-1.html#text"},{"title":"5 Theory","offset":16323,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"5.1 Forward Direction: LeJEPA Learns the World Model","offset":16926,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Proof Sketch.","offset":16984,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Interpretation.","offset":17776,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"5.2 Converse Direction: The Gaussian is Unique","offset":18270,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Proof Sketch.","offset":18968,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"5.3 Approximate Identifiability","offset":19624,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Interpretation.","offset":19941,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"5.4 Optimal Latent Planning","offset":20422,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Interpretation.","offset":20645,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"6 Experiments","offset":20942,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"6.1 Forward: Linear Identifiability from Gaussian Latent Variables","offset":21276,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Scaling to High Dimensions.","offset":22667,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"6.2 Converse: Non-Gaussian Latent Variables Break Linear Identifiability","offset":23474,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Latent-Distribution Sweep.","offset":23651,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Pixel-Based RL Trajectories.","offset":24101,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"6.3 Approximation Bound and Loss Predict Identifiability","offset":25143,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"6.4 Linear Identifiability Enables Latent-Space Planning","offset":25946,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"7 Limitations","offset":26876,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Are the latents Gaussian?","offset":26894,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"What if the dimension is wrong?","offset":27490,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Finite samples and optimization.","offset":27964,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"8 Discussion","offset":28414,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Outlook.","offset":28431,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Practical Implications.","offset":28959,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"A Theoretical Foundation for World Models.","offset":29847,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Acknowledgments","offset":30444,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Appendix Overview","offset":30713,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-2.html#text"},{"title":"Appendix A Proof of Theorem 5.1 (Forward Direction)","offset":33705,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-3.html#text"},{"title":"A.1 Intuition: Why Should This Force Linearity?","offset":33761,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-3.html#text"},{"title":"A.2 Hermite Polynomial Background","offset":35671,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-3.html#text"},{"title":"Definitions and orthogonality.","offset":35911,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-3.html#text"},{"title":"Multivariate extension.","offset":37646,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-3.html#text"},{"title":"Hermite expansion of measure-preserving maps.","offset":38836,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-3.html#text"},{"title":"A.3 Proof","offset":40077,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-3.html#text"},{"title":"Setup and notation.","offset":40092,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-3.html#text"},{"title":"Step 1: The transition attenuates each Hermite degree by ρ d \\rho^{d} .","offset":41940,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-3.html#text"},{"title":"Step 2: Different Hermite components do not interact across views.","offset":45966,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-3.html#text"},{"title":"Step 3: Nonlinearity strictly reduces correlation.","offset":48442,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-4.html#text"},{"title":"Step 4: Assembly.","offset":51841,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-4.html#text"},{"title":"Step 5: Recovery of the transition.","offset":52648,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-4.html#text"},{"title":"Appendix B Proof of Theorem 5.2 (Gaussian Uniqueness)","offset":53378,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-4.html#text"},{"title":"B.1 Background","offset":53650,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-4.html#text"},{"title":"B.2 Proof","offset":56691,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-4.html#text"},{"title":"Appendix C Proof of Theorem 5.3 (Approximate Identifiability)","offset":57978,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-4.html#text"},{"title":"Setup and notation.","offset":58293,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-4.html#text"},{"title":"C.1 Step 1: Alignment gap controls nonlinear energy","offset":59856,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-4.html#text"},{"title":"C.2 Step 2: Whitening constraint forces M M near orthogonal","offset":63386,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"C.3 Step 3: Pythagorean decomposition","offset":64966,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"Role of the Gaussian on the latent side vs. the output side.","offset":66348,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"Appendix D Proof and Discussion of Theorem 5.4","offset":67107,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"Proof.","offset":67158,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"D.1 Worked Applications","offset":68480,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"Goal-reaching.","offset":68613,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"Linear-quadratic regulation.","offset":69294,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"General rotation-invariant objectives.","offset":71445,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"D.2 Scope: encoder versus action-conditioned dynamics","offset":72295,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"D.3 Connection to classical optimal control","offset":73887,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"Appendix E Alternative Proof via Dirichlet Energy","offset":74963,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"E.1 Setup","offset":75382,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"E.2 Proof","offset":75975,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"Step 1: Reduction to Dirichlet energy.","offset":75990,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"Step 2: The expected log-determinant is zero.","offset":78271,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-5.html#text"},{"title":"Step 3: Lower bound via AM-GM and Jensen.","offset":79073,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html#text"},{"title":"Step 4: Equality forces orthogonal Jacobian.","offset":80633,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html#text"},{"title":"Step 5: Orthogonal Jacobian implies linearity (Mazur–Ulam).","offset":81388,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html#text"},{"title":"Generalization.","offset":83288,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html#text"},{"title":"E.3 Comparison of the Two Proofs","offset":83833,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html#text"},{"title":"Appendix F Prior Work: Connection to Slow Feature Analysis","offset":84624,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html#text"},{"title":"SFA and its optimization problem.","offset":85137,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html#text"},{"title":"Nonlinear identifiability via SFA.","offset":86129,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html#text"},{"title":"Eigenfunction structure and Sturm–Liouville theory.","offset":86860,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html#text"},{"title":"The Gaussian specialization and Hermite polynomials.","offset":88137,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html#text"},{"title":"Sequential versus simultaneous extraction.","offset":89233,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html#text"},{"title":"The role of isotropic transitions.","offset":90586,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html#text"},{"title":"Beyond Gaussian latents.","offset":92381,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html#text"},{"title":"Relation to Mehler’s formula and maximal correlation.","offset":93692,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-6.html#text"},{"title":"Practical differences: learned features versus fixed kernels.","offset":94605,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"Connections to diffusion maps and spectral geometry.","offset":95588,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"Summary of the relationship.","offset":96581,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"Appendix G Lean Verification","offset":97335,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"Scope and methodology.","offset":97856,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"G.1 What Is Axiomatized and Why","offset":98712,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"1. Hermite polynomial infrastructure.","offset":98801,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"2. Sturm–Liouville spectral theory.","offset":99249,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"3. Standard analysis results requiring API software engineering.","offset":99854,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"4. Measure-theoretic and matrix-analytic facts.","offset":100320,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"5. Trajectory pushforward for expected costs.","offset":100783,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"What is verified.","offset":101687,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"G.2 Build Information","offset":102464,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"Appendix H Experimental Details and Additional Results","offset":102690,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"H.1 Nonlinear Mixing Functions","offset":102749,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"H.2 Network Architecture and Training","offset":104728,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"H.3 Training Procedure","offset":105251,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"H.4 Hyperparameters","offset":106309,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"H.5 Evaluation Metrics","offset":107273,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"H.6 Grid Search","offset":108499,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-7.html#text"},{"title":"H.7 Latent Variable Distribution Sweep Across Mixings","offset":110400,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-8.html#text"},{"title":"H.8 Approximate Bound Decomposition","offset":111976,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-8.html#text"},{"title":"H.9 Loss Predictivity","offset":112886,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-8.html#text"},{"title":"H.10 Scaling Experiment","offset":113322,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-8.html#text"},{"title":"H.11 Reacher Experiment","offset":116513,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-8.html#text"},{"title":"Environment and rendering.","offset":116542,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-8.html#text"},{"title":"Encoder architecture.","offset":118477,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-8.html#text"},{"title":"Training.","offset":119440,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-8.html#text"},{"title":"Evaluation.","offset":120090,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-8.html#text"},{"title":"Latent distributions.","offset":120752,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-8.html#text"},{"title":"Gaussianity and autocorrelation jointly determine identifiability.","offset":121806,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-8.html#text"},{"title":"Additional results on OU and Trajectory conditions.","offset":124494,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-8.html#text"},{"title":"H.12 Planning Experiment","offset":126095,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-9.html#text"},{"title":"Setup.","offset":126506,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-9.html#text"},{"title":"Metric.","offset":127314,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-9.html#text"},{"title":"Interpretation.","offset":127851,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-9.html#text"},{"title":"Instructions for reporting errors","offset":128145,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/part-9.html#text"}],"text_code_points":129109,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f140b8fcc13ea085ce6ad170e090380c8d373416f791960778c4b216f6e5c8f6/source.md"}