Pith. sign in

REVIEW 2 major objections 4 minor 49 references

Analytic Planning under Uncertainty with Moment Closure

T0 review · 2 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Paper shows Bellman backups can be computed in closed form, freeing model-based RL from Monte Carlo target variance.

desk verdict The stress-test concern lands: the moment-matched collapse is an avoidable, untested approximation, but the compatibility principle and Eq. (7) hold up — send it out and demand the exact-mixture ablation. read the letter →

arxiv 2608.02519 v1 pith:4OFRTUHK submitted 2026-08-03 cs.LG cs.AI

Shishir Sharma , Doina Precup This is my paper
classification cs.LGcs.AI
keywords model-basedreinforcementlearninganalyticBellmanbackupmomentclosureGaussiantransitionmodelradialbasisfunctionuncertaintypropagationdeepensemblescontinuouscontrol
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper pursues a yes/no question: can model-based reinforcement learning propagate full predictive distributions through the Bellman backup without Monte Carlo sampling and without restricting policy or reward structure? It argues yes, and proves it by decomposing the Q-function so that the greedy maximization disappears, leaving an expectation of the value function alone; a compatibility principle then makes that expectation analytic. Concretely, with a Gaussian transition model and a value function written as a mixture of radial basis functions, the next-state value expectation has a closed form that carries both predictive mean and covariance. The upshot, according to the paper, is a planning update with no sampling variance and well-calibrated uncertainty, which outperforms sampling-based and certainty-equivalent baselines under stochastic observations.

What carries the argument

The central object is the moment-closed Bellman expectation — the identity E_{s'~N(mu,Sigma)}[sum_i w_i exp(-1/2 ||s'-c_i||^2_{Lambda_i})] = sum_i w_i |I+Sigma Lambda_i|^{-1/2} exp(-1/2 ||mu-c_i||^2_{Lambda_i(I+Sigma Lambda_i)^{-1}}). The work this identity does is to make the value expectation a deterministic function of the predictive mean and covariance, so uncertainty propagates exactly through the backup. Two structural choices carry the argument: the bounded quadratic action-value parameterization, which turns max_a Q(s',a') into V(s') and thus frees the policy from any compatibility requirement, and the moment-compatibility design principle itself, which generalizes the Gaussian–RBF p

What would settle it

Run the algorithm in a domain with bimodal transition dynamics (for example, a control task where a door either opens or stays shut), keep the moment-matched Gaussian target, and compare action rankings against a high-sample Monte Carlo target; if the analytic ranking diverges from the Monte Carlo ranking in a way that degrades control performance, the central claim that moment closure preserves uncertainty-aware planning is contradicted.

Watch

Extended reading notes

Core claim

Under a quadratic action-value parameterization (the bounded NAF construction), the Bellman backup reduces to an expectation over next states of the state-value function alone, because the argmax over actions is given in closed form. The paper's key discovery is a compatibility principle: if the predictive transition distribution and the value representation are chosen so that this expectation is a known function of the distribution's moments, the backup is analytic. With a heteroscedastic Gaussian dynamics model (trained by maximum likelihood, with an ensemble condensed by moment matching) and a radial-basis value function, the expectation is Eq. (7): sum_i w_i |I+Sigma Lambda_i|^{-1/2} exp

Load-bearing premise

The load-bearing premise is that the next state, given a state and action, is adequately represented by a single moment-matched Gaussian; if the true transition is multimodal or strongly non-Gaussian, the analytic expectation in Eq. (7) is exact for the wrong distribution and there is no sampling error to reveal the mistake.

Editorial extensions

If this is right

  • Bellman targets become sampling-free: the analytic expectation replaces Monte Carlo averaging, eliminating target variance entirely.
  • Both aleatoric noise (via the learned covariance) and epistemic uncertainty (via ensemble moment matching) enter the target exactly through the predictive covariance, so planning explicitly accounts for them.
  • Greedy action selection stays closed-form under bounded action sets, so uncertainty propagation does not constrain the policy class.
  • Empirically, the method remains stable as observation noise increases whereas sampling-based (SAC-style) and certainty-equivalent (mean-only) baselines degrade.
  • Because the expectation is closed form, the target is a deterministic function of state and action, which can reduce update variance in the critic regression.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct extension the paper leaves open: apply the same moment-closure in a learned latent state space, where the RBF value class is compact, to scale beyond the low-dimensional control domains tested here.
  • The method assumes the predictive distribution is unimodal Gaussian after moment matching; in environments with genuinely multimodal transitions (e.g., a mixture of outcomes), the closed form is exact only for the wrong distribution, so its robustness in such regimes is an open empirical question.
  • The closed-form expectation is differentiable with respect to the predictive mean and covariance, which suggests a possible route to gradient-based policy improvement or to analytic, uncertainty-aware exploration bonuses — implications the paper does not explore.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper proposes MoCA, a model-based RL algorithm that constructs Bellman targets analytically rather than by sampling. Using a bounded NAF/BNAF quadratic action-value parameterization, the inner maximization is removed from the Bellman backup, leaving E_{s'|s,a}[V(s')]. The paper pairs a Gaussian dynamics model with an RBF-mixture value function, yielding the closed-form expectation in Eq. (7). An ensemble of Gaussian dynamics models is collapsed to a single moment-matched Gaussian (Eqs. (4)-(5)), and the resulting analytic target is used in a TD3-style update. Experiments on Cartpole and Pendulum with Gaussian observation noise claim improved robustness, lower target variance, and well-calibrated predictive uncertainty compared with SAC, BNAF, certainty-equivalent, and Monte Carlo variants.

Significance. If the result held as stated, the paper would offer a valuable sampling-free Bellman backup for a useful class of continuous-control problems. The derivation of Eq. (7) is correct, and the NAF/BNAF decoupling of greedy selection from expectation is a clean and sound idea. The central 'moment-compatibility' principle is plausible. However, the specific moment-matching approximation is unnecessary for tractability and discards exactly the ensemble information the paper claims to preserve. The empirical support is narrow: two low-dimensional tasks, all with Gaussian observation noise, so the ensemble mixture is never stressed. These issues undermine the central claim as currently formulated, though the framework appears fixable.

major comments (2)
  1. [§3.4, Eq. (4)–(5); Algorithm 1 lines 10–11] The ensemble mixture is collapsed to a single moment-matched Gaussian, but this is not required for analyticity. Since V is an RBF mixture (Eq. (6)), the exact ensemble expectation is also closed form: E_{hat p}[V(s')] = (1/K) sum_k sum_i w_i |I + Sigma_k Lambda_i|^{-1/2} exp(-1/2 ||mu_k - c_i||^2_{A_{i,k}}), with A_{i,k} = Lambda_i(I + Sigma_k Lambda_i)^{-1}. This is just Eq. (7) applied to each member, costing O(K M d^3) instead of O(M d^3) and requiring no additional model forward passes. Thus Eq. (5) integrates a different distribution without a tractability gain, discarding multimodality and inter-member disagreement beyond second moments. The experiments (§5.2) use deterministic dynamics plus Gaussian observation noise, so the mixture is near-Gaussian and the approximation is never stressed. Replace the moment-matched target with the exact mixture expectation or provide an ablation
  2. [§5.2–§5.3] Empirical support is narrow: only Cartpole and Pendulum, both low-dimensional, and all injected stochasticity is Gaussian observation noise with deterministic underlying dynamics. This keeps the ensemble mixture near-Gaussian and never probes non-Gaussian or multimodal predictive distributions, which is exactly where moment matching changes the target. Moreover, Gaussian observation noise on nonlinear dynamics does not induce a Gaussian transition distribution in the observed-state MDP, so the protocol does not actually isolate well-specified Gaussian transitions. The '1σ coverage' diagnostic (Fig. 3, middle) calibrates the learned dynamics model, not the value or planning uncertainty; it does not substantiate the paper's 'well-calibrated predictive uncertainty' claim for control. Report results with non-Gaussian/multimodal transition noise and higher-dimensional tasks, and include the e
minor comments (4)
  1. [§5.3] Specify the sampling distribution for MoCA-MC: does it sample from the moment-matched Gaussian N(mu_bar, Sigma_bar) or from the ensemble mixture? The text says MoCA-MC computes 'the same expectation' as MoCA, but MoCA already uses an approximation, so this needs to be unambiguous.
  2. [Fig. 3] Define '1σ coverage' and 'epistemic std' operationally in the caption or text. As shown, it is unclear whether coverage refers to the dynamics model's predictive distribution or to the value estimate, which matters for the calibration claim.
  3. [§4, Algorithm 1 line 8] Clarify the bootstrap resampling: each dynamics member is trained on an independent bootstrap resample of the minibatch B, but the resample size and replacement mechanism are not specified.
  4. [§3.3, Eq. (3)] Specify whether the projection Pi_A is with respect to the Euclidean norm or the P-weighted norm. The equality max_{a in A} Q(s,a) = V(s) requires the appropriate projection, or a proof for box constraints with non-diagonal P.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Eq. (7) is a parameter-free Gaussian/RBF identity, the NAF reduction is an explicit parameterization, and no load-bearing self-citations are present.

full rationale

Eq. (7) is derived in Appendix B.2 by completing the square; it is an exact identity for any Gaussian N(mu,Sigma) and RBF basis, with no fitted parameters and not containing the Bellman target as an input. The reduction of the Bellman backup to E[V(s')] follows from the quadratic action-value parameterization (2)/(3), where V(s) is defined as max_a Q(s,a); this is a structural construction, not a prediction smuggled from data. The Gaussian/RBF pairing is introduced as a 'moment-compatibility' design choice, and the paper explicitly labels the moment-matched ensemble collapse as an approximation, citing Lakshminarayanan et al. (2017) as external prior work; the existence of an exact mixture alternative is a modeling/efficiency remark, not a circular reduction. No citation to the present authors' prior work is load-bearing; Plaksin & Martyanov and Gu et al. are independent external results used as stated. The 1-sigma coverage diagnostic is an internal self-consistency check rather than an external prediction, but it is not used to derive Eq. (7) or the algorithm, so at most a minor reporting caveat. The central closed-form target is therefore self-contained under its stated Gaussian assumptions.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central math rests on the Gaussian integral and on NAF/BNAF structure; the empirical claim rests on the Gaussianity/moment-matching assumption and on hand-set RBF support. No new physical or conceptual entities are introduced.

free parameters (5)
  • RBF centers c_i = Hand-set per environment (App. B.4)
    Centers are fixed uniformly over practical bounds; the value function and the analytic backup depend on this coverage.
  • RBF lengthscales lambda_i = Cartpole [0.6,0.5,0.5,3,4]; Pendulum [0.5,0.5,2]
    Hand-chosen per dimension (App. B.5); controls smoothness of V and the sensitivity of the closed-form expectation to Sigma.
  • Number of RBF bases M = 256 (Cartpole), 128 (Pendulum)
    Set by dimensionality and perceived value complexity (App. B.3); no automatic selection criterion is given.
  • Curvature keep-alive penalty (lambda, p_min) = lambda=1e-4; p_min=0.5 (Cartpole), 2.0 (Pendulum)
    Hand-tuned penalty in App. C affecting all BNAF-based variants; could influence the relative performance of the baselines.
  • Training hyperparameters (LRs, exploration schedule, target period, batch) = See App. D
    Standard hyperparameters but hand-set; they affect the learning curves and comparisons.
assumptions (6)
  • standard math The expectation of a Gaussian RBF under a Gaussian density is given by Eq. (7), including determinant and precision-shrinkage terms.
    Used in Appendix B.2; computed by completing the square; no empirical input.
  • domain assumption The learned/true conditional p(s'|s,a) is Gaussian with mean mu_theta and covariance Sigma_theta (Section 3.4).
    Underlies Eq. (5) and hence all analytic targets; the experiments inject Gaussian observation noise to try to satisfy it, but nonlinear true dynamics make exact Gaussianity unlikely.
  • domain assumption Moment-matching the ensemble mixture to a single Gaussian preserves E[V(s')] accurately enough for planning (Eq. 4-5).
    This is the weakest load-bearing approximation; if higher moments or multimodality matter, the analytic backup is exact only for the wrong distribution. No diagnostic is reported.
  • domain assumption The bounded NAF (BNAF) action-value family can approximate the optimal Q in the test MDPs with bounded suboptimality (Section 3.3, citing Plaksin and Martyanov 2022).
    The reduction from max_a Q to V depends on this expressivity claim; the paper assumes the cited applicability conditions hold for Cartpole/Pendulum without verifying control-affine/quadratic-cost structure.
  • domain assumption Fixed RBF centers and lengthscales provide sufficient coverage and smoothness for V to represent the optimal value in the region visited by the policy (Appendix B.4-B.5).
    The value representation is not learned for centers or bandwidths; if hand-set support is poor, the value approximation and the analytic expectation both degrade.
  • domain assumption Off-policy TD learning with nonlinear function approximation and bootstrapped targets remains stable enough for the method's performance claims (Algorithm 1).
    No convergence guarantee is given; the paper relies on standard RL practice and twin-critic heuristics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Analytic Planning under Uncertainty with Moment Closure." pith.science (2026). https://pith.science/paper/4OFRTUHK

@misc{pith2026260802519,
  author       = {Pith},
  title        = {Pith review of: Analytic Planning under Uncertainty with Moment Closure},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4OFRTUHK}},
  note         = {Machine review of arXiv:2608.02519}
}
read the original abstract

Effective model-based reinforcement learning in stochastic environments requires planning that accounts for predictive uncertainty. Propagating full state distributions analytically offers a principled way to do this, but has traditionally required restrictive policy or reward structures to remain tractable. Consequently, modern deep reinforcement learning has largely retreated to either stochastic sampling, which introduces significant target variance, or deterministic point estimates that ignore predictive covariance entirely. We investigate whether distribution-aware planning is possible without these constraints. Using a quadratic action-value parameterization, we first reduce the Bellman backup to an expectation over the state-value function alone; the key idea is then a compatibility principle between the predictive transition distribution and the value function class, under which this expectation is analytic in the distribution's moments. We instantiate this principle with a Gaussian transition model paired with a radial-basis value function, yielding a closed-form backup that propagates both predictive mean and covariance. Empirically, our approach reduces target variance and yields well-calibrated predictive uncertainty under stochastic observations in continuous control, providing a principled framework for planning with learned distribution models.

Figures

Figures reproduced from arXiv: 2608.02519 by the authors.

Figure 1
Figure 1. Analytic versus Monte Carlo estimation of the Bellman expectation E[V (s ′ )], on a one-dimensional toy example with a Gaussian predictive distribution p(s ′ ) and a radial-basis value function V (s ′ ). (A) The expectation is the integral of the density-weighted value p(s ′ ) V (s ′ ) (shaded); Here equal to E[V (s ′ )] = 0.057. (B) A Monte Carlo estimate from K samples is unbiased but is high-variance, especially … view at source ↗
Figure 2
Figure 2. Learning curves on Cartpole under Gaussian observation noise. Top row: homoscedastic noise with standard [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Return and uncertainty diagnostics on Pendulum under Gaussian observation noise ( [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 6 linked inside Pith

  1. [1]

    2018 , publisher=

    Reinforcement learning: An introduction , author=. 2018 , publisher=

    2018

  2. [2]

    2018 , eprint=

    Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models , author=. 2018 , eprint=

    2018

  3. [3]

    2017 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Information theoretic MPC for model-based reinforcement learning , author=. 2017 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2017 , organization=

    2017

  4. [4]

    , author=

    Gaussian Processes in Reinforcement Learning. , author=. NIPS , volume=. 2003 , organization=

    2003

  5. [5]

    Proceedings of the 28th International Conference on machine learning (ICML-11) , pages=

    PILCO: A model-based and data-efficient approach to policy search , author=. Proceedings of the 28th International Conference on machine learning (ICML-11) , pages=. 2011 , organization=

    2011

  6. [6]

    arXiv preprint arXiv:1312.5602 , year=

    Playing atari with deep reinforcement learning , author=. arXiv preprint arXiv:1312.5602 , year=

    arXiv

  7. [7]

    Machine learning , volume=

    Q-learning , author=. Machine learning , volume=. 1992 , publisher=

    1992

  8. [8]

    2017 , eprint=

    Neural Network Dynamics for Model-Based Deep Reinforcement Learning with Model-Free Fine-Tuning , author=. 2017 , eprint=

    2017

Show all 49 references
  1. [9]

    nature , volume=

    Human-level control through deep reinforcement learning , author=. nature , volume=. 2015 , publisher=

    2015

  2. [10]

    Advances in neural information processing systems , volume=

    Action-conditional video prediction using deep networks in atari games , author=. Advances in neural information processing systems , volume=

  3. [11]

    arXiv preprint arXiv:1611.07078 , year=

    A deep learning approach for joint video frame and reward prediction in atari games , author=. arXiv preprint arXiv:1611.07078 , year=

    arXiv

  4. [12]

    arXiv preprint arXiv:1205.4839 , year=

    Off-policy actor-critic , author=. arXiv preprint arXiv:1205.4839 , year=

    arXiv

  5. [13]

    nature , volume=

    Mastering the game of Go with deep neural networks and tree search , author=. nature , volume=. 2016 , publisher=

    2016

  6. [14]

    arXiv preprint arXiv:1509.02971 , year=

    Continuous control with deep reinforcement learning , author=. arXiv preprint arXiv:1509.02971 , year=

    arXiv

  7. [15]

    arXiv preprint arXiv:1707.06347 , year=

    Proximal policy optimization algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

    arXiv

  8. [16]

    2018 , howpublished =

    Joshua Achiam , title =. 2018 , howpublished =

    2018

  9. [17]

    Advances in Neural Information Processing Systems , year =

    Michael Janner and Justin Fu and Marvin Zhang and Sergey Levine , title =. Advances in Neural Information Processing Systems , year =

  10. [18]

    2019 , eprint=

    Learning Latent Dynamics for Planning from Pixels , author=. 2019 , eprint=

    2019

  11. [19]

    2020 , eprint=

    Dream to Control: Learning Behaviors by Latent Imagination , author=. 2020 , eprint=

    2020

  12. [20]

    2017 , eprint=

    Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles , author=. 2017 , eprint=

    2017

  13. [21]

    2025 , eprint=

    Deterministic Uncertainty Propagation for Improved Model-Based Offline Reinforcement Learning , author=. 2025 , eprint=

    2025

  14. [22]

    Proceedings of the 28th International Conference on International Conference on Machine Learning , pages =

    Deisenroth, Marc Peter and Rasmussen, Carl Edward , title =. Proceedings of the 28th International Conference on International Conference on Machine Learning , pages =. 2011 , isbn =

    2011

  15. [23]

    2016 , eprint=

    Continuous Deep Q-Learning with Model-based Acceleration , author=. 2016 , eprint=

    2016

  16. [24]

    International conference on machine learning , pages=

    Continuous deep q-learning with model-based acceleration , author=. International conference on machine learning , pages=. 2016 , organization=

    2016

  17. [25]

    2020 , eprint=

    Planning with Expectation Models , author=. 2020 , eprint=

    2020

  18. [26]

    2019 , eprint=

    Continuous control with deep reinforcement learning , author=. 2019 , eprint=

    2019

  19. [27]

    2018 , eprint=

    Addressing Function Approximation Error in Actor-Critic Methods , author=. 2018 , eprint=

    2018

  20. [28]

    2018 , eprint=

    Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor , author=. 2018 , eprint=

    2018

  21. [29]

    2024 , eprint=

    Mastering Diverse Domains through World Models , author=. 2024 , eprint=

    2024

  22. [30]

    2022 , eprint=

    Temporal Difference Learning for Model Predictive Control , author=. 2022 , eprint=

    2022

  23. [31]

    Proceedings of the International Conference on Artificial Intelligence and Statistics (AISTATS) , year =

    Model-Based Uncertainty in Value Functions , author =. Proceedings of the International Conference on Artificial Intelligence and Statistics (AISTATS) , year =

  24. [32]

    Neural Information Processing Systems , year =

    Deterministic Uncertainty Propagation for Improved Model-Based Offline Reinforcement Learning , author =. Neural Information Processing Systems , year =

  25. [33]

    2007 , publisher=

    Optimal control: linear quadratic methods , author=. 2007 , publisher=

    2007

  26. [34]

    Proceedings of the 2005, American control conference, 2005

    A generalized iterative LQG method for locally-optimal feedback control of constrained nonlinear stochastic systems , author=. Proceedings of the 2005, American control conference, 2005. , pages=. 2005 , organization=

    2005

  27. [35]

    2012 IEEE/RSJ International Conference on Intelligent Robots and Systems , pages=

    Synthesis and stabilization of complex behaviors through online trajectory optimization , author=. 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems , pages=. 2012 , organization=

    2012

  28. [36]

    Advances in Neural Information Processing Systems , volume=

    Continuous deep Q-learning in optimal control problems: Normalized advantage functions analysis , author=. Advances in Neural Information Processing Systems , volume=

  29. [37]

    2025 , eprint=

    MuJoCo Playground , author=. 2025 , eprint=

    2025

  30. [38]

    2018 , eprint=

    DeepMind Control Suite , author=. 2018 , eprint=

    2018

  31. [39]

    arXiv preprint arXiv:1812.05905 , year=

    Soft actor-critic algorithms and applications , author=. arXiv preprint arXiv:1812.05905 , year=

    arXiv

  32. [40]

    International conference on machine learning , pages=

    Addressing function approximation error in actor-critic methods , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  33. [41]

    Advances in Neural Information Processing Systems , volume=

    Deterministic uncertainty propagation for improved model-based offline reinforcement learning , author=. Advances in Neural Information Processing Systems , volume=

  34. [42]

    International conference on machine learning , pages=

    Learning latent dynamics for planning from pixels , author=. International conference on machine learning , pages=. 2019 , organization=

    2019

  35. [43]

    International Conference on Learning Representations , year=

    Dream to Control: Learning Behaviors by Latent Imagination , author=. International Conference on Learning Representations , year=

  36. [44]

    Lillicrap , title=

    Danijar Hafner and Jurgis Pasukonis and Jimmy Ba and Timothy P. Lillicrap , title=. CoRR , volume=. 2023 , cdate=

    2023

  37. [45]

    International Conference on Machine Learning, PMLR , year=

    Temporal Difference Learning for Model Predictive Control , author=. International Conference on Machine Learning, PMLR , year=

  38. [46]

    Advances in neural information processing systems , volume=

    Simple and scalable predictive uncertainty estimation using deep ensembles , author=. Advances in neural information processing systems , volume=

  39. [47]

    Proceedings of the 28th International Joint Conference on Artificial Intelligence , pages=

    Planning with expectation models , author=. Proceedings of the 28th International Joint Conference on Artificial Intelligence , pages=

  40. [48]

    Nature , year =

    Hafner, Danijar and Pasukonis, Jurgis and Ba, Jimmy and Lillicrap, Timothy , title =. Nature , year =. doi:10.1038/s41586-025-08744-2 , url =

    doi

  41. [49]

    Advances in neural information processing systems , volume=

    What uncertainties do we need in bayesian deep learning for computer vision? , author=. Advances in neural information processing systems , volume=

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.