acceptodds
Under review as a conference paper at ICLR 2027

The Connectivity Cost of Shared Coordinates in World Models

Abstract

In world models, each action can be implemented with a concise connectivity pattern under its own optimal coordinate system, but the optimal coordinate systems for different actions are generally inconsistent. World models require all actions to share the same coordinate system in order to read and update a consistent representation, yet this causes an increase in the number of connections required for the actions. Existing work has primarily focused on coordinate alignment methods that reduce the number of connections, but a specific characterization of how the connectivity scale changes when actions move from their respective coordinates to shared coordinates is still lacking. This paper investigates this connectivity expansion induced by shared state coordinates. In the symmetric linear model analyzed in this paper, we prove that for families of actions lacking common structure, each action requires only a linear number of connections under its own optimal coordinates, whereas the minimum number of connections typically grows to quadratic order under shared coordinates. The reason is that the connectivity structures required by different actions generally cannot be sufficiently aligned and reused within a single coordinate system, and this gap persists even when approximate compression is allowed. We further show that if different actions share a fixed sparse connectivity pattern under the same coordinate system, the number of connections required for the shared state can still remain at a linear scale, and we present a structure that avoids this increase in asymptotic order.

Primary area:
learning theory
Keywords:
world models, shared representations, sparse connectivity, representation locality

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UErcOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies how many state-to-state connections are needed when several action-conditioned dynamics must use the same coordinates. It compares the minimum block connectivity under separate and shared orthogonal bases. For generic symmetric operators, the paper gives a linear-to-quadratic separation and studies whether this gap persists under approximate compression. It also shows that a common sparse interaction structure can keep the shared representation sparse. The experiments cover synthetic matrix families and local Jacobians from world models trained on an NRI spring system.

Critical strengths

The paper gives a concrete formulation of the difference between individually simple dynamics and a simple shared representation. The finite-budget analysis is an important part of this formulation: it asks whether the difference remains after allowing weak connections to be discarded, rather than relying only on exact sparsity.

The learned-model experiments make a useful effort to isolate the effect of shared interaction support. Reporting prediction error and separate-basis compression error alongside the shared-coordinate penalty helps distinguish the effect from differences in model fit or individual-action compressibility. The slot-model comparison also provides a relevant control.

Critical weaknesses
  1. The connection between the main scaling result and learned world models remains limited. The linear-to-quadratic separation holds for generic symmetric operators. For general real operators, both costs are already quadratic, and the learned Jacobians in Section 6.4 are mostly nonsymmetric. The paper acknowledges this distinction and includes a potential-model control in Appendix E.4. However, I still find the link between the scaling results and the measured compression penalty in the learned models insufficiently developed. The experiments show that support sharing matters, but do not establish how the theoretical connectivity counts explain the size of the observed penalty.
  2. It is unclear what the measured penalty tells us about a world model beyond its local compressibility. The main learned-model result is a difference in FbF_b, computed from local residual Jacobians in a small spring-system setting. This supports the proposed structural interpretation, but the paper does not show whether the difference matters when sparse dynamics are actually used for prediction or rollout, or whether it translates into a computational benefit. A new architecture or a SOTA comparison is not necessary here. My concern is that the paper stops short of explaining what a world-model researcher should conclude or do differently after observing a large FbF_b. This limits the significance of the contribution for me.
Questions for authors
  1. How should the general-real connectivity result in Eq. (10) be connected to FbF_b, which measures approximation error under a per-action budget? Does the theory give a prediction about the dependence of this measured penalty on the budget or state dimension?

  2. Is there evidence in the existing experiments or analysis that the difference in FbF_b matters when the compressed dynamics are used for prediction or rollout? More generally, what conclusion should a model designer draw from a large penalty when the original models have similar prediction errors?

  3. The paper distinguishes its objective from joint diagonalization and joint block diagonalization. Could the authors make the comparison more precise for Theorems 1 and 2, identifying the closest existing results and what additional result is needed to obtain the connectivity count and finite-budget bound?

3 comments
AuthorsOpenReview ↗

General Real Operators and the Finite-Budget Penalty F_b

We agree that the linear-to-quadratic separation in the original manuscript was established for generic symmetric operators and should not be directly applied to learned Jacobians, which are typically nonsymmetric. However, the central question of this paper is more general. Different actions may each have simple local dynamics, but this does not mean that they can share a single orthogonal coordinate frame without additional cost. From the perspective of operator geometry, the issue is not merely how many matrix entries are nonzero. It is whether the local subspace decompositions induced by different actions are compatible with one global coordinate frame.

  1. For general real operators, we have further derived the exact connection counts. For even block width,

    C_{\mathrm{sep}}=k\frac{m(m+1)}{2}, \qquad C_{\mathrm{sh}}=(k-1)m^2+\frac{m(m+1)}{2}.

Thus, although both schemes have quadratic complexity for general real operators, shared coordinates still incur a quadratic absolute overhead in the number of connections:

C_{\mathrm{sh}}-C_{\mathrm{sep}}
=\frac{(k-1)m(m-1)}{2}.

The ratio of the two complexities approaches 2-1/k. This shows that nonsymmetry changes the relative growth of the sharing cost but does not remove the structural cost of requiring different actions to use the same coordinate frame. We have also derived corresponding upper and lower bounds for odd block widths.

  1. We have also established a theoretical connection between the exact connection counts and the finite-budget penalty F_b. Let E_B^{\mathrm{total}} denote the minimum error when all actions share one orthogonal basis under a total connection budget of B. Then

    E_{kb}^{\mathrm{total}}\leq E_b^{\mathrm{sh}},

which gives

\boxed{
F_b=E_b^{\mathrm{sh}}-E_b^{\mathrm{sep}}
\geq E_{kb}^{\mathrm{total}}-E_b^{\mathrm{sep}}.
}

This relation directly connects the theoretical approximation obstruction under a total budget to the shared penalty defined per action in our experiments. Furthermore, for general real Gaussian operator families, within an explicitly specified quadratic budget range, separate coordinates permit exact representation, whereas shared coordinates retain, with high probability, an approximation-error lower bound that does not vanish with dimension. In 15 numerical instances with k=4,r=2, the shared scheme still incurs feasible relative errors of approximately 5.42%–6.82%, even after more extensive optimization checks.

These results characterize the structural obstruction to shared representations from two perspectives. First, the geometric degrees of freedom in a common coordinate frame are insufficient to eliminate all cross-subspace couplings for a generic operator family. Second, this obstruction is not limited to exact sparsity and may persist under finite-budget approximation. We also distinguish theoretical guarantees from experiments on learned models. The Gaussian asymptotic result does not directly predict the specific value of F_b for each learned model. Rather, it provides a further theoretical characterization of the obstruction to shared representations for general nonsymmetric operator families. We will clarify the applicability conditions of these results and their relationship to the experiments with learned Jacobians.

AuthorsOpenReview ↗

Practical Significance of F_b for Prediction, Rollout, and Efficiency

Interpreting F_b first requires distinguishing the compressibility of an individual action from the representational compatibility of multiple actions. A world model must maintain a persistent internal state, so it cannot freely redefine its state coordinates whenever the action changes. Even if each action admits a simple local representation on its own, all actions need not remain simple in one common set of state coordinates. F_b measures the additional structural cost imposed by this cross-action sharing constraint, rather than the predictive quality of the original model itself.

To test whether this structural difference affects prediction in practice, we added three real-robot tasks: Lift, Can, and Tool Hang. We use the same frozen dynamics model and compress only its residual update, thereby isolating the effect of the shared-coordinate constraint. The improvements in real-trajectory prediction MSE achieved by Separate over Shared are as follows:

  • Lift: 2.67% for one-step prediction and 12.23% for ten-step prediction.
  • Can: 4.22% for one-step prediction and 13.87% for ten-step prediction.
  • Tool Hang: 3.00% for one-step prediction and 10.99% for ten-step prediction.

These results indicate that the structural cost of shared coordinates is not reflected only in the number of nonzero connections in local Jacobians. This incompatibility can also affect practical prediction when the compressed dynamics are actually used to update robot states.

Shared coordinates retain complexity in every dynamics update. By contrast, action-specific coordinates allow each action to maintain a simpler local structure, but require coordinate conversion when the action changes. Appendix D quantifies this cost and gives the average connectivity cost associated with action switches. Accordingly, action-specific coordinates may offer computational advantages when action switches are infrequent, whereas coordinate conversion itself may become a major overhead when switches are frequent.

For ten-step prediction, both sparse schemes proposed in the paper achieve approximately 1.32–1.33x speedup over the full dense model. This demonstrates computational savings at the level of the overall task.

These findings point to more than the fact that Separate outperforms Shared in some experiments. They suggest a broader model-design principle. When F_b is large, the difficulty may not lie in compressing any individual action, but in making different actions share the same local state representation economically. Model designers can respond by increasing the shared connection budget, exploiting a common sparse interaction structure, revisiting the state partition, or using action-conditioned coordinates when the cost of switching is acceptable. When the sharing penalty is small, a unified persistent state representation may be the more economical choice.

AuthorsOpenReview ↗

Relation to Joint Diagonalization and Joint Orthogonal Sparsification

Thank you for recommending relevant work on matrix decompositions. We will expand the comparison with these studies and clarify the contribution of our work. Joint diagonalization, finest common block-diagonal decomposition, and the identification of common block structures under congruence transformations mainly address the existence and identification of shared diagonal structures in a family of matrices. However, the absence of nontrivial common invariant subspaces does not imply that several matrices cannot remain sparse in the same basis. For example, a diagonal matrix with distinct diagonal entries and an irreducible symmetric tridiagonal matrix have no nontrivial common invariant coordinate blocks, yet both have only a linear number of nonzero entries in their original coordinates. Therefore, the finest common block-diagonal decomposition alone cannot determine the number of connections under arbitrary sparse supports.

The work by Ba et al. on joint orthogonal Hessian sparsification is also closely related to ours. It studies how multiple Hessian matrices can be sparsified through a common orthogonal transformation and analyzes the corresponding optimization problem and its properties. We further investigate two quantitative questions underlying this problem:

  1. For a generic operator family, what are the exact minimum connection counts under action-specific coordinates and shared coordinates? In other words, what is the additional cost of enforcing a common orthogonal basis?

  2. Does this sharing cost persist when weak connections are allowed to be discarded?

We therefore do not propose a new joint-diagonalization or joint-sparsification algorithm. Instead, starting from the geometric compatibility of common representations, we quantify the structural cost that multiple actions must incur when they share the same state coordinates, and characterize the conditions under which this cost persists under finite-budget approximation. We will discuss the closest related studies more precisely and clarify the additional theoretical guarantees provided by our two theorems.

Reviewer_Q3F3OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper considers the following stetting. We are given kk d×dd \times d matrices A1,A2,…,AkA_1, A_2, \dots, A_k, with d=mrd = mr. These matrices represent kk different actions on the full system, whose state space is Rd=Rmr\mathbb{R}^{d} = \mathbb{R}^{mr}. The full state space RmrR^{mr} is seen as a direct sum of mm copies of Rr\mathbb{R}^{r}, with the idea being that each of the mm copies denotes a different "part" of the full system. The matrices can then naturally be seen as m×mm \times m block matrices where each block is of size r×rr \times r. The ijijth block of the matrix AlA_{l} then determines the "interaction" between the iith and jjth "parts" of the system in the action represented by the matrix AlA_l.

The paper is concerned with the problem of choosing a basis of the full state space Rd\mathbb{R}^d so as to minimize the number of non-trivial such interactions: in particular, with the problem of minimizing the total number of non-zero blocks in the kk matrices.

Most of the paper is concerned with the case when the matrices are symmetric real matrices, and the basis changes considered are thus orthogonal matrices. One of the main results, Theorem 1, is that if the same orthogonal matrix must be used for each of the AiA_i, then, in the "generic" case, the minimum achievable total number of non-zero blocks is θ(m2)\theta(m^2) (it is easy to see that the minimum value in this case is at most (k−1)m2+m(k-1)m^2 + m, which can be obtained by choosing a matrix that diagonalizes one of the AiA_i, and the paper shows that in the "generic" case, one cannot do better than (k−1)m2+1(k-1)m^{2} + 1, by a parameter counting argument). The paper compares this with the case when a different matrix can be used for each of the AiA_i, in which case each of them can be diagonalized to get a total of at most km=Θ(m)km = \Theta(m) non-zero blocks, and again by a parameter counting argument, the paper shows that this cannot be improved in the "generic" case. Thus there is a gap between the θ(km2)\theta(km^2) non-zero blocks in the case of having a "shared" basis transformation vs a θ(km)\theta(km) non-zero blocks in the case of having a different orthogonal basis transformation for each "action" AiA_i.

The paper then explores this result more quantitatively in two ways:

(1) By looking at a notion of normalized distance from structures with bounded number of non-zero blocks (eq. 7 in the paper)

(2) By looking at structures where which blocks are allowed to be non-zero is fixed (Section 5).

The results for (2) appear to be essentially definitional. The result for (1) is that when the total number of non-zero blocks is below (k−1)m2(k-1)m^2, then, when the matrices AiA_i are chosen from an appropriate independent Gaussian ensemble, the normalized distance is at least a "constant" (that depends upon rr, kk, and ρ\rho, where ρ\rho is roughly the ratio of the allowed-number of non-zero blocks with km2km^2) with high probability (exponentially small in mm). This result is formalized using a standard ϵ\epsilon-net argument in the appendix.

In the introduction, the paper phrases all these results in terms of relevance of world models, with the mm copies seen as different parts of the same system. The questions asked in this paper therefore concern how to create sparse representations of actions by an appropriate choice of coordinates.

Critical strengths

The paper formulates a simple linear algebraic question that is of potential relevance to world models.

Critical weaknesses

As stated above, the paper frames the introduction in terms of world models. However, the first two sections contain almost no forward references to the technical results to discuss in detail how the quantitative results fit in the context of world models. In general, the paper is also very hard to read, with some of the readability concerns described later.

On the technical side, it is not clear to me that Theorem 1, the main result, would be new to any expert on world models. The upper bounds on number of non-zero blocks would be immediate to anyone with a linear algebra background, and the fact that these are almost tight in the the generic setting is also not hard (a proof using more "standard" tools is outlined below in the "Questions"). Further, many of the results, such as Proposition 1 and Theorems 3 and 4, are essentially definitional: they are trivial consequences of the definitions.

Further, the paper consider only orthogonal basis transformations, and notes that if the matrices are not symmetric, then the main result of the θ(m)\theta(m) vs. θ(m2)\theta(m^2) separation between a "shared" and "independent" basis transformations disappears (lines 228-237).

Readability and presentation problems

Several parts of the paper are written in way that are repetitive or hard to read/understand. A related issue is that results deferred to the appendix sometimes seem to show something different from what is claimed in the main text. Some examples of such concerns are below.

  • line 147, paragraph beginning "The central gap lies elsewhere." The sentences following this first sentence seem to simply repeat the arguments of the preceding paragraph. (In particular, this is the argument summarized in the sentence "The question that matters is how many additional connections shared coordinates require, compared with letting each action use its own optimal coordinates." from the preceding paragraph.)

  • lines 259-260. "Below the critical density, discarded relations carry a non-vanishing fraction of the operator energy with high probability, so approximate sparsity does not remove the obstruction." I am not sure how the conclusion of this sentence follows, and I am even at a loss as to what the sentence formally means.

  • line 237 says "Appendix A treats odd block widths and the asymptotics in r." But the relevant parts of Appendix A only seem to say "For odd r a complex pair can be forced to straddle two block boundaries, so the separate count can exceed k m(m + 1)/2 by a term of order m; the regime conclusion is unaffected, but the exact equality should be stated for even r only." This does not quite look a treatment of the case of odd block widths.

  • At various points, new terms are introduced, without definition, to refer to notions that were being discussed earlier in the paper under different names.

    • The word "obstruction" seems to occur only thrice in the paper: once in the title of Theorem 2, once here, and once on line 270 ("The obstruction results raise the next question."), so it is not clear what exactly is being referred to as an obstruction now and why. Similarly, the word "envelope" appears for the first time at the beginning of Section
    • Similarly, the word "envelope" appears at the beginning of Section 5 without any definition, where the simpler "block structure" (which also seems to be used, for example in Theorem 3) is used again.
    • In another instance, the paper suddenly starts to refer to some of the matrices AiA_i as "Jacobians" on line 229, without any further explanation (which maps are these Jacobians of?).
  • At some points, simple to describe objects are stated in a needlessly convoluted way. For example, on line 244-246 which described a "Frobenius-isotropic Gaussian on $\mathrm{Sym}_d(\mathbb{R})". The description given here leaves open the question of whether such a distribution even exists. A much shorter and cleaner description would be to include the standard construction of such a distribution.

  • Some of the phrasing in the appendix seems as if it is from personal research notes rather than for a final paper. In addition to the example from line 237 above, there is, for example, on lines 1015-1026, which says, "The constant degrades as log⁡(1/c)=O(1/δ)\log(1/c) = O(1/\delta) near the threshold, so any report of cc must state δ\delta".

Questions for authors

See also the weaknesses section.

  • l. 168. "A block therefore represents one directed relation from one state unit to another." This sentence probably belongs a bit later, when block matrices are discussed. The representation in the previous display is just a standard direct sum/"concatenation" decomposition of a vector space. It seems that "block" here refers to block matrices

  • l. 214. The Theorem claims in eq. (9) Csh(A)=(k−1)m2+mC_{sh}(\mathcal{A}) = (k-1)m^2 + m. But the "dimension counting" proof in the appendix only shows the lower bound Csh(A)≥(k−1)m2+1C_{sh}(\mathcal{A}) \ge (k-1)m^2 + 1 (eq. 29 on line 819 in the appendix). In particular, the lower and upper bounds in eqs. (29) and (30) do not match unlike what is claimed in the statement of the theorem.

    Also, is Sard's theorem really needed for the proof? It seems that the dimension counting can be done just using the "area theorem", which would imply that since the map ΦS\Phi_{\mathcal{S}} is nice, the Hausdorff dimension of its image can be at most the Hausdorff dimension of the domain. As the dimension counting argument shows, this is less than the ambient dimension of the co-domain. Hence, the image must have smaller than full Hausdorff dimension, and hence its full-dimensional Lebesgue measure must be zero.

  • l. 229. "Learned Jacobians are not symmetric." So far, no Jacobians had been discussed in the paper. Should "Jacobians" be "actions" here? If these are Jacobians, what function are they Jacobians of?

  • l. 252. "Low-budget shared representations form a closed cone and EB is the squared normalized distance to it." Isn't it more accurate to say that they form a finite union of closed cones? If you take a non-negative linear combination of two such representations with different patterns but the same "budget", the budget of the resulting patter might be more than either of the starting representations.

  • Proposition 1. This is very confusingly worded. The paper says "block envelope contains all m diagonal blocks and the two directed blocks associated with each edge". This seems to suggest there are m+2m + 2 non-zero blocks in an action. But then in the next sentence, the paper says this number is m+2∣E∣m + 2|E|. Then, in the statement of Proposition 1, the paper uses the notion of "graph envelope" which has not been defined. (What is probably meant is that each action is now allowed only diagonal blocks and the blocks corresponding to the edges of the graph: it would have been much clearer if it were actually stated this way).

  • Theorem 3 follows directly from eq. (16) and (7), so I am not sure why a proof sketch is given.

    • Similarly for Theorem 4, which, as stated is true only for action families of size kk,
  • l.353. This paragraph seems out of place here. It probably belongs in Section 6 on Experiments. Also, it is clear from the definitions so far that the quantities do not depend upon the "gauge" (which is a slightly confusing term to use in this case, since here it just seems to mean the choice of a basis), so this paragraph seems quite confusing.

6 comments
AuthorsOpenReview ↗

Technical Clarifications and Proof Revisions

Thank you for identifying issues with the logical flow of the proofs, the definitions, and the presentation. We have revised the paper accordingly and added the necessary explanations to improve its clarity and readability. We have also strengthened the critical-budget result. By removing the off-diagonal block pairs with the smallest energies, we obtain a uniform O(1/m) error upper bound that holds for every nonzero symmetric operator family. Thus, although a generic family cannot be represented exactly for finite m, its optimal approximation error can still approach zero as m increases.

  1. We have reordered the explanation that a block represents a directed relation and the definition of the block matrix. We now clarify that block (i,j) describes the influence of state unit j on state unit i.

  2. The previous coarse lower bound was insufficient to establish the exact formula. In the complete proof in Appendix A.3, we remove the redundant counting caused by within-block rotations. This gives a lower bound matching the constructive upper bound and establishes the exact formula in Theorem 1.

  3. We adopted your suggested dimension argument. For each fixed support, after removing the redundancy from within-block rotations, the parameter space has dimension below the ambient space whenever the budget is below the claimed minimum. The smooth representation map therefore has a measure-zero image. Taking the finite union over supports gives the lower bound, without using Sard's theorem.

  4. We now specify that the Jacobian is the derivative of the action-conditioned state transition f_a, describing how a small change in the current state affects the next state. Our experiments analyze Df_a - I, namely the component remaining after the identity update is subtracted.

  5. We now explicitly describe the low-budget set as a finite union of closed cones, generally nonconvex. Scaling a feasible representation preserves feasibility, whereas adding representations with different supports may exceed the budget. The proof uses closedness and scaling invariance, not closure under addition.

  6. The graph envelope specifies the connection positions available to all actions, and each action may use only a subset of them. Each action uses at most m + 2|E| blocks, giving at most k(m + 2|E|) blocks across k actions. We now distinguish the allowed positions from the positions that are actually nonzero, as well as per-action counts from counts over the entire family.

  7. The error bound in Theorem 3 follows by retaining the common sparse component. We have shortened the proof to this construction. The result identifies a structural condition under which a low-budget shared representation has small approximation error.

  8. Theorem 4 is similarly direct. We have shortened its proof and clarified that the result applies to any nonzero symmetric action family with k >= 1.

  9. We have moved the coordinate-calibration discussion to the experimental protocol. The exact optimized quantities are invariant under a common orthogonal change of coordinates, whereas general invertible changes can alter the energy metric. We now state that exact whitening makes linearly related coordinates orthogonally equivalent when they use the same calibration distribution with positive definite covariance.

AuthorsOpenReview ↗

Relevance to World Models

Thank you for pointing out that the original manuscript did not sufficiently connect the mathematical results to the specific issues they address in world models. We will provide a more detailed discussion of this connection.

The central issue in our work is not whether the dynamics of an individual action can be compressed. Rather, it is whether the locally simple dynamics of different actions can remain simple in the same coordinates when a world model must maintain a persistent internal state representation. These are two distinct levels of the problem. For each action, we may find coordinates suited to its local dynamics. However, when the model must continue to use the same state representation as the action changes, these local coordinate systems can no longer be chosen independently. We study this cross-action representational compatibility and the structural cost it incurs.

Our theoretical results address different aspects of this issue. Theorem 1 quantifies the exact connectivity cost imposed by shared coordinates. Theorem 2 further examines whether this cost persists when approximate compression is allowed and weak connections can be discarded. Section 5 shows that if different actions have a common sparse interaction structure, they may still admit a simple shared representation even without sharing the same eigenstructure. Appendix D analyzes action-specific coordinates from another perspective. Although they can reduce the connectivity cost of each dynamics update, they introduce coordinate-conversion costs when actions switch, which may become the dominant overhead.

Our study therefore does not simply compare shared and separate representations to determine which is better. Instead, it addresses a broader model-design question. When the local dynamics of different actions are structurally incompatible, should the representation complexity be paid as additional connections under shared coordinates, or shifted to coordinate conversions between actions? We also identify the structural conditions under which this additional cost can be avoided.

We have also added real-robot prediction experiments to test whether this structural difference affects practical use. Using the same frozen dynamics model and 200 real-robot demonstrations from Tool Hang, the separate representation reduces the one-step and ten-step real-state prediction MSE by approximately 3.00% and 10.99%, respectively, compared with the shared representation. This indicates that the local structural cost caused by cross-action representation constraints can translate into task-level gains when the compressed dynamics are actually used for state prediction. The purpose of this metric is to provide a diagnostic dimension for model design that may previously have been overlooked. In addition to checking whether each action can be compressed independently, one should examine whether all actions can remain simple together in a single persistent state coordinate system, and make the choice in light of prediction quality and coordinate-conversion costs.

We will make these connections more explicit in the first two sections so that the persistent-state requirement in world models, the common-coordinate problem in the theory, and the prediction results in the experiments are clearly linked.

AuthorsOpenReview ↗

Substantive Contributions Beyond Standard Linear Algebra

Orthogonal diagonalization and the other tools used in this paper are indeed classical. The main contribution, however, lies in identifying a common-representation problem that has not been quantitatively characterized in sufficient detail, and in extending its analysis from exact structure to approximation under a finite budget.

First, we do not merely ask whether several matrices can be jointly diagonalized. We characterize the minimum number of connections required by multiple operators under specified conditions, as well as the additional cost of using shared coordinates. This differs in an important way from joint diagonalization. Joint block decomposition concerns whether several operators have common invariant subspaces. However, the absence of nontrivial common invariant subspaces does not imply that the operators cannot share a sparse representation. For example, a diagonal matrix with distinct diagonal entries and an irreducible symmetric tridiagonal matrix may have no nontrivial common invariant coordinate blocks, yet both have only a linear number of nonzero entries in their original coordinates. Thus, showing that no common block decomposition exists is not sufficient to establish a lower bound on the number of connections under arbitrary sparse supports.

From the perspective of matrix geometry, genuinely distinct block partitions can be viewed as the quotient space

\mathcal F_{m,r}=O(d)/O(r)^m

. After removing the rotational freedoms within individual blocks, the remaining degrees of freedom are precisely those that affect the connectivity between different state subspaces. The lower bound in Theorem 1 is based on the dimension relationship between these effective degrees of freedom and the operator support space. This geometric perspective allows us to rigorously characterize the generic optimal number of connections over all shared coordinates and all admissible support patterns, rather than merely providing a construction for one particular common basis.

Second, after allowing all possible shared orthogonal bases and low-budget support patterns, how far can the original operator family be from the entire set of low-budget representations? Through covering-number and high-dimensional probability analyses, Theorem 2 establishes a positive approximation-error lower bound that holds simultaneously over all shared bases and low-budget supports, under a specified independent Gaussian operator model and budget range. This advances the question from identifying operator families that cannot be represented exactly to determining whether the difficulty persists, with a stable error cost, after weak connections are discarded.

Together, these two results establish the main mathematical relationship pursued in this paper. Using classical tools, we provide a continuous quantitative characterization that runs from the geometric degrees of freedom of common representations, through exact connectivity complexity, to finite-budget approximation barriers. We further explain the relevance of this characterization to action-conditioned world models.

Reviewer_LrwhOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper looks at the impact of learning actions in world models. In modeling languages (such as PDDL), actions are typical sparse, only impacting a small fraction of the state representation. When learning world models, because there is no direct regularization between actions, learning the basis for a single action is not likely to be a good basis for another action. The paper attempts to characterize the cost of shared representations.

Critical strengths

The paper formalizes questions around models and learned state representations. Different regimes are explored in reference models and results match the theoretical predictions. In particular, in some regimes using a common sparse envelope maintains concise connectivity, but this isn’t guaranteed in general.

Critical weaknesses

This paper does not interact with recent work on using graph Laplacians for environmental dynamics. One such example is Shehmar et al, 2026 (ICML) https://arxiv.org/pdf/2602.05031 It seems there is some overlap in that using the graph Laplacian may avoid many of the difficulties being described/analyzed in this paper.

Figure 1 has a mix of fonts and is mis-sized with respect to the rest of the paper. This looks like something that belongs on a poster versus something informative within a paper. The terminology used in the caption is not defined. One might guess mm is the number of edges, but it isn’t clear what $k is in the figure. This figure needs improvement.

Questions for authors

Can you comment on work using graph Laplacians?

1 comment
AuthorsOpenReview ↗

Graph Laplacians and Figure 1 Revisions

Q1 — Graph Laplacians and the Scope of Our Work

We revisited Shehmar et al.’s ALPS and agree that it is a relevant study to discuss and compare in the related-work section. ALPS represents environment states and their transitions as a graph, and uses the eigenvectors of the graph Laplacian to construct state representations that capture environmental reachability and timescales. These representations are used for high-level subgoal discovery and planning, while its low-level forward model still predicts dynamics in the original state space. ALPS therefore focuses primarily on spectral representations of the environment’s state space, whereas our work studies the action-conditioned coupling structure among different state units within the same state. Although both works involve spectral structure and state representations, the mathematical objects under study are different.

We also agree that Laplacian structure can avoid the connectivity costs discussed in our paper in certain cases. For example, if all action operators are functions of the same symmetric Laplacian, they can share its eigenbasis. Likewise, if all actions have a common physical sparse interaction graph, a common sparse envelope can preserve linear connectivity. We will add this related work and clarify that common spectral structure and common sparse structure are the compatible cases identified by our theory. We further investigate how much connectivity cost shared coordinates incur when these structural conditions do not hold, and whether that cost remains under approximate compression.

Q2/Q3 — Figure 1: Typography and Definitions

Thank you for your comments on the readability of the figure and for suggesting improvements. In the original Figure 1, we used Gabriola and Times New Roman to distinguish the research problem from the specific findings. However, as you pointed out, the inconsistent fonts, undefined terms, and mismatch with the typography and scale of the rest of the paper made the figure difficult to read. In the revised Figure 1, we have unified the fonts and explicitly defined the terms used in the figure. We have also adjusted the text size to ensure that it is proportionate to the main text after compilation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.