betaFeedback

WorksComputer Science

When Does Less Information Pay? An Exact Finite Atlas of Representation, Payoff, and Cost

Abstract

A representation that retains one bit about a four-state world can solve a task perfectly, while another that retains 1.5 bits on average reaches only 75% accuracy. The difference lies in which distinctions each representation preserves. We examine this relationship in a uniform four-state world by enumerating all admissible representations and binary action rules. Constructions are compared both by replacing one component while holding the others fixed and by choosing the best action rule separately for each module. We consider three tasks, their mixtures, and four separate hypothetical cost tariffs. In some comparisons, the representation retaining less world information has the higher payoff; in others, the more informative representation does. There are also pairs with different information amounts but equal payoff–that is, equal task accuracy. If a representation preserves all existing distinctions and adds new ones, its best zero-cost payoff cannot decrease. Lower-information advantages therefore depend on which distinctions are retained. Equal counts of opposing advantages under representation replacement follow from coding symmetry. In broader comparisons, the shares depend on how repeated realizations are counted. A different task set or resource price can change which member of a fixed pair performs better. These results are exact, complete censuses of the specified finite sets of constructions and comparisons. The atlas provides a reproducible worked model for examining how action constraints, counting rules, and cost assumptions affect comparisons. It illustrates established distinctions between information quantity and decision value rather than proposing a new general law of their relation. I(W;R)I(W;R) measures retained world information, and payoff measures task accuracy.

Introduction

If an agent cannot retain all the available information about its environment, which distinctions between states should it preserve? The answer depends on the task: a representation may contain more information about the world overall, yet fail to preserve the particular distinctions needed for the right action.

Consider four equiprobable world states, labelled by two-digit binary codes: 00, 01, 10, 11. The agent’s task is to identify the first, left-hand bit. A representation that distinguishes 00, 01 from 10, 11 allows error-free answers and retains one bit of world information. Another representation merges 00 with 10 but distinguishes the other two states. It retains 1.5 bits on average, yet cannot always answer correctly.

In this example, one bit of information means that two of the four equiprobable possibilities remain. This happens every time with the first representation. With the second, half the time—in states 01 or 11—the world state is known exactly, providing two bits of information. The other half of the time, the possibilities are 00 and 10, providing one bit. The average is (1+2)/2=1.5(1 + 2)/2 = 1.5 bits. The fractional value describes the representation’s average informativeness.

The first representation’s advantage comes from which distinctions it preserves, not from the reduction in information itself. As shown in § 3.2, simply removing distinctions from a given representation cannot improve its best attainable accuracy.

The distinction between information quantity and task relevance is already central to information-theoretic compression. The deterministic information bottleneck balances representation entropy against information retained about a target ([13]). Tegmark and Wu [15] study such trade-offs for binary classification. These criteria concern information about a target, whereas our payoff measures the accuracy of a chosen action rule; the two are not interchangeable.

Tan et al. [14] search the space of deterministic groupings for a Pareto frontier: representations for which neither compression nor target information can be improved without sacrificing the other. Their method also applies to other two-objective clustering problems, and they use exhaustive enumeration as a small-case reference. Our purpose is different: to retain all specified pairwise comparisons, including nonoptimal constructions, and examine the consequences of fixed versus separately optimized action rules, repeated realizations, task mixtures, and distinct costs. The atlas is not a new algorithm for finding optimal representations.

The motivation for this study came from Donald Hoffman’s The Case Against Reality and the evolutionary modelling literature it discusses ([4, 5, 9]). A central question in this debate is whether fitness-oriented strategies can outperform strategies described as veridical. Addressing it requires distinguishing representation, information, payoff, and biological fitness. A representation specifies which world states an agent distinguishes. Information quantifies statistical dependence between the representation and the world. Payoff scores the success of selected actions in a given task. Relating this payoff to reproductive success requires a separate model.

Previous studies show why the answer cannot be separated from the comparison conditions. Advantages for simpler or interface-like strategies have been found under particular payoff and cost assumptions ([5]); a general mathematical result has also been obtained for specified strategy classes ([9]). Other factors may matter, however: information that is useful only when several input signals are combined ([6]), environmental change ([1, 8]), and the need to solve multiple independent tasks using one representation that does not adjust to the current goal ([2]). Preserving relations between world states is a further, distinct criterion: information quantity alone is insufficient to assess it ([10]).

Our initial aim was to reproduce selected early evolutionary simulations. Within the search described in Supplementary Section S1, we did not identify a single executable package sufficient to reproduce those calculations exactly. This finding does not extend to all related literature: Berke et al. [2], for example, published code and data subsequently used by Mattes [8]. In the present study, we investigate our own finite model rather than reproduce the early simulations.

We ask a narrower question: **when does a representation that retains less information about the world allow a given task to be solved more accurately? Separately, we examine how a charge for a chosen resource changes net performance—the payoff after subtracting costs.

The model is limited to four equiprobable world states and three binary tasks. In each task, one answer corresponds to two states and the other to the remaining two. Within these bounds, we completely enumerate deterministic representations and action rules—mappings without random choice.

We make two kinds of comparison. In the first, one component of an agent’s construction changes while the others remain fixed. For example, the grouping of world states can change while the response rule is retained. This comparison shows the result of a particular replacement under specified constraints. In the second, we determine the best accuracy available to each representation by choosing its response rule separately. Here, we compare capabilities after optimization.

The counting rule also matters. One representation may belong to several constructions with different sensory filters or different amounts of spare capacity. We therefore compare two censuses: one includes all such realizations, and the other counts each representation once.

We then examine how the advantage within a fixed pair depends on the task or task mixture being evaluated, and on the resource charge under each of four specified tariffs. All censuses retain cases favoring the less informative or the more informative member, equal information, and equal payoffs.

The contribution is the explicit, exactly computed set of comparisons and its reproducible implementation. It brings the effects of action constraints, counting rules, task mixtures, and four cost assumptions into one inspectable finite model. The underlying distinctions between information and decision value are established in earlier theory, including the Blackwell comparison of signals [3, 11], and information-cost trade-offs [7]. We use these principles to explain the census, not claim them as new discoveries.

Model and methods

World, tasks, and payoff

Let the ordered world be

W=(00,01,10,11),P(w)=1/4,W = (00, 01, 10, 11), \qquad P(w) = 1/4,

with binary action set A={0,1}A = \{0, 1\}. The three balanced task partitions are

yfirst=(0,0,1,1),ysecond=(0,1,0,1),yxor=(0,1,1,0).y_{\mathrm{first}} = (0, 0, 1, 1), \qquad y_{\mathrm{second}} = (0, 1, 0, 1), \qquad y_{\mathrm{xor}} = (0, 1, 1, 0).

The first task asks for the left bit, the second for the right bit, and exclusive OR (XOR) asks whether the bits differ. The first two have archived identifiers bit_0 and bit_1. These are all balanced binary task partitions up to output complement: there are (42)=6\binom{4}{2} = 6 ways to assign label 1 to two states, and exchanging labels 0 and 1 leaves three partitions.

A correct action scores 1 and an incorrect action scores 0:

ut(w,a)=1[a=yt(w)].u_t(w, a) = \mathbf{1}[a = y_t(w)].

Payoff is the average score across the four world states: the proportion of correct answers. For example, an agent that answers correctly in three of the four states has a payoff of 3/4=0.753/4 = 0.75, corresponding to 75% correct answers. A rule with full access to the world can score 1 on each task; a constant action scores 1/21/2.

Figure 1 follows the calculation from world state to representation, action, and payoff. A separately specified cost may then be subtracted.

Figure 1. The quantities used in the model: retained information, task payoff, mechanism cost, and net performance after cost.

Representations and action rules

A representation R=r(W)R = r(W) groups world states into categories. The rule rr is deterministic: the same state always belongs to the same category. We denote the number of categories used by KRK_R.

To give identical groupings identical encodings, we read the states in the specified order. The first category encountered receives label 0, the next new category receives label 1, and so on. The labels used form the set Zr={0,…,KR−1}Z_r = \{0, \ldots, K_R - 1\}.

For example, 0011 means that the first two states belong to one category and the last two to another. This encoding is called a restricted-growth string (RGS). Renaming categories does not create a new representation.

The four world states can be partitioned into categories in 15 ways: one uses one category, seven use two, six use three, and one uses four.

A finer representation preserves every distinction made by a coarser one and may subdivide its categories further. If at least one category is actually split, the refinement is strict. Such representations are nested. If neither of two representations preserves all the distinctions made by the other, they are non-nested.

An action rule, or policy, assigns each category the answer 0 or 1. We denote it by π:Zr→{0,1}\pi: \mathcal{Z}_r \to\{0,1\}. A representation with KRK_R categories has 2KR2^{K_R} such rules, and we enumerate them all.

Across all 15 representations, this gives 94 representation-policy combinations. Each combination is evaluated on three tasks, yielding 94×3=28294 \times3 = 282 result rows.

The payoff of a given policy and the best payoff available to a representation are defined as follows:

Vt(r,π)=14∑w∈W1[π(r(w))=yt(w)],Vt∗(r)=max⁡πVt(r,π).V_t(r,\pi) = \frac{1}{4}\sum_{w \in W} \mathbf{1}[\pi(r(w)) = y_t(w)], \qquad V_t^*(r) = \max_{\pi} V_t(r,\pi).

The payoff ceiling Vt∗(r)V_t^*(r) is the highest average score attainable on task tt with the given representation by choosing the best policy. If several policies achieve this maximum score, we retain them all.

Mutual information I(W;R)I(W;R) measures how much knowing the representation category reduces uncertainty about the world state on average [12]. We call this quantity retained world information.

Uncertainty is measured by entropy HH, expressed in bits. For four equally likely states, H(W)=log⁡24=2H(W) = \log_2 4 = 2 bits. Because the world state uniquely determines the representation category, retained world information equals the entropy of the representation itself:

I(W;R)=H(R).I(W;R) = H(R).

However, information about the world does not necessarily help solve a particular task. We therefore also measure I(Yt;R)I(Y_t;R): how much information about the correct answer YtY_t is contained in the representation.

Finally, the policy maps the representation to an action AoutA_{out}. The quantity I(Yt;Aout)I(Y_t;A_{out}) measures how much information about the correct answer is contained in the selected action. We thus distinguish information available in the representation from information reflected in the agent’s response.

To test how a policy performs on another task, we first find all policies that are optimal for the source task. We then evaluate each policy, without changing it, on the other two tasks. For each source–target task pair, we retain the lowest and highest payoffs among the source-optimal policies. This shows what the previous response rules achieve without adjustment.

For 15 representations, three source tasks, and two possible target tasks, this gives 15×3×2=9015 \times3 \times2 = 90 result rows.

Figure 2 brings together all 15 representations. It shows which states they group together, how much information they retain, and what accuracy they can achieve. Here the policy is optimized separately for each task.

Figure 2. All 15 ways to partition the four world states into categories. Within a row, equal letters and fills indicate states in the same category. RGS is the grouping code; KRK_R is the number of categories used; II is retained world information in bits. The orange bars show the best accuracy on each task after choosing an optimal policy. For example, representation 0011 allows the first, left-hand bit of the state to be identified without error.

Modules and coded constructions

Each agent construction specifies which world states the agent can distinguish, which distinctions it retains in its representation, and how it chooses an action.

The sensory filter FF determines which distinctions between world states are available to the agent. States grouped together by the filter reach the agent as the same input.

The representation RR determines which of these available distinctions are retained. It may further merge inputs into shared categories, but cannot recover distinctions unavailable through the filter. For example, if the filter cannot distinguish states 00 and 01, the representation must also assign them to the same category.

Provisioned capacity K∈{1,2,3,4}K \in\{1,2,3,4\} limits the number of representation categories. Not all of this capacity needs to be used: with K=4K = 4, for example, a representation may contain only two categories. We therefore distinguish provisioned capacity KK from the realized category count KR=∣R∣K_R = |R|. Every admissible construction satisfies KR≤KK_R \le K.

It remains to specify which action the agent chooses for each category. The decision code DD assigns it an answer of 0 or 1. The action code α\alpha either leaves this answer unchanged or reverses it: 0 becomes 1, and 1 becomes 0.

The complete construction thus has five components:

g=(F,K,R,D,α).g = (F, K, R, D, \alpha).

For world state ww, the representation determines a category, the decision code assigns it an answer, and the action code turns that answer into the final action:

Ag(w)=α(D(R(w))).A_g(w) = \alpha(D(R(w))).

We call the combination of filter, capacity, and representation (F,K,R)(F, K, R) a module. It does not yet include an action rule. The module payoff ceiling is the best score under a given objective across all admissible combinations of DD, α\alpha. We assess how well the module can perform a task; learning or evolutionary search for a suitable action rule is not modelled here.

With capacity K=1K = 1, there are 15 admissible modules; with K=2K = 2, 47; with K=3K = 3, 59; and with K=4K = 4, 60. In total, this gives 181 modules.

The decision code assigns each representation category one of two answers: 0 or 1. For ∣R∣|R| categories, this gives 2∣R∣2^{|R|} choices of decision code. Each can be combined with two action codes, either keeping or reversing the answer. Summed over all modules, these choices give 1,424 complete coded constructions. The detailed count is provided in Supplementary Section S3A.

We use model genotype as a label for a complete construction with its chosen decision and action codes. A model phenotype comprises the module and the action the construction assigns to each world state. Different combinations of codes can produce the same actions. When the module is also the same, these constructions have different genotypes but the same phenotype. These terms describe the construction’s design and behaviour, without biological assumptions.

Different encodings are counted separately, even when they specify the same actions. When determining the payoff ceiling, we also retain all code combinations that achieve the maximum for the given module and objective.

Single-task and multi-task objectives

We evaluate each construction on individual tasks and on combinations of tasks. For multiple tasks, the overall score is the arithmetic mean of the payoffs. For example, for a pair of tasks, we add the two payoffs and divide by two. All tasks within the chosen set receive equal weight.

The three tasks form seven nonempty sets: three single tasks, three distinct pairs, and one set containing all three tasks. Each set defines a scoring objective — a rule for calculating the overall score.

Let SS denote the chosen task set and ∣S∣|S| the number of tasks in it. If Vt(g)V_t(g) is the payoff of construction gg on task tt, its overall score is

VS(g)=1∣S∣∑t∈SVt(g).V_S(g) = \frac{1}{|S|} \sum_{t \in S} V_t(g).

The same construction gg, with an unchanged action rule, is evaluated on every task in the set. The term mixture refers to averaging scores across a fixed set of tasks, not to a sequence of environmental changes. Mixtures use exactly the answer labelings in § 2.1. Independently inverting one task’s answers is not included as an additional mixture variant: with a shared action rule, it can change the result. Completeness up to answer inversion refers to the partitions of individual tasks.

Which pairs are compared?

Constructions can be compared in different ways. The first question is: what changes when only one component is replaced and the rest remain fixed? The second is: which of two modules can achieve a higher payoff when the best action rule is chosen separately for each? We consider both cases.

In the one-component comparison, two constructions differ in exactly one of the components FF, KK, RR, DD, α\alpha. For example, two representations RR can be compared with the same FF, KK, DD, α\alpha. Both constructions must satisfy the model’s constraints.

Replacing only RR preserves the number of categories: the unchanged decision code DD contains one answer per category. Retained information can still differ. For example, two categories can divide the four states into groups of sizes 3 + 1 or 2 + 2. These comparisons change the grouping at a fixed category count, rather than removing categories from a representation.

For each fixed setting of the other four components, we compare every pair of admissible alternatives. Order within a pair does not matter: comparing A with B and B with A counts as one comparison. A replacement need not correspond to a single mutation; mutation distances and transition rates are not specified here.

For the information/payoff census, only pairs that replace the representation RR are used from this group.

In the all-module-pairs comparison, we compare any two distinct modules from the complete set. For each module and each objective, the decision and action codes that maximize payoff are chosen separately. These maxima are then compared. This census shows differences in the modules’ capabilities after optimization, rather than the effect of replacing one component.

We also check the effect of counting a representation more than once. The same representation can belong to several modules with different filters FF or provisioned capacities KK. Some representations therefore appear more often than others in the all-module-pairs comparison.

In this counting control, modules with the same representation are collapsed into a single entry. We then compare all pairs of the 15 distinct representations. The control uses the same saved results and changes only how they are counted, without adding constructions. It is termed a quotient census in the supplement.

It is important to distinguish a pair from its evaluations. A comparison identity is one pair of constructions or modules. A comparison row evaluates that pair under one objective. Each pair is evaluated under all seven objectives, so there are seven times as many rows as pairs.

CensusNumber of pairsNumber of rows
One-component replacements, all five types9,56866,976
Of these: representation RR replacements only1,4169,912
All pairs of the 181 modules16,290114,030
All pairs of the 15 representations, without repetitions105735

Table 2.

The number of pairs among nn distinct objects is n(n−1)/2n(n-1)/2: each pair is counted once.

Figure 3 shows the two comparison methods and the counting control. Each census has its own set of pairs and total number of comparisons, so the resulting fractions are reported separately.

Figure 3. Part A compares constructions that differ in one component. Part B compares the best attainable payoffs of all module pairs; one representation may belong to several modules. Part C shows the counting control, in which these repeated realizations are collapsed and pairs of distinct representations are compared.

Together, these comparisons show under which conditions lower or higher retained world information is associated with higher payoff: for a specific representation replacement or after choosing the best action rule for each module. The representation-once control allows us to determine how much the fractions of such cases depend on some representations belonging to more modules than others.

Information and payoff orders

We now compare two properties of the members of each pair: how much world information they retain and what payoff they receive. We ask which member has the higher payoff — the one retaining less or more information.

We compare information amounts using exact values of I(W;R)I(W; R). For each objective, we assign the comparison result to one of four groups:

  1. Equal information amounts. Payoffs may be either equal or different.

  2. Less information, higher payoff. The member retaining less world information receives a higher payoff than the other.

  3. More information, higher payoff. The member retaining more world information receives a higher payoff than the other.

  4. Different information amounts, equal payoffs. The difference in retained information is not accompanied by a payoff difference under this objective.

Each result row belongs to exactly one of these four groups. Equal-information comparisons are kept separate: neither member can be identified as more or less informative in these comparisons.

Equal information amounts do not imply identical representations. Values of I(W;R)I(W; R) can coincide even when the representations retain different distinctions between world states.

Changes of order between objectives

We next ask whether one member of a pair can receive a higher payoff on an individual task, while the other receives a higher payoff on a task mixture.

To compare the results, we assign the members the labels ”left” and ”right”. These labels remain fixed across all objectives. The payoff difference is

ΔV=Vleft−Vright.\Delta V = V_{left} - V_{right}.

If ΔV>0\Delta V > 0, the left member has the higher payoff; if ΔV<0\Delta V < 0, the right member does. At ΔV=0\Delta V = 0, the payoffs are equal.

We record a single-task-versus-multi-task reversal if one member strictly outperforms the other on at least one single task, and the other member strictly outperforms it on at least one mixture. In other words, the payoff differences must have opposite nonzero signs. Equal payoffs alone do not count as a reversal.

For module pairs, we also run a broader check: does each member have a strict advantage under at least one of the seven objectives? Here the opposing results may concern, for example, two single tasks, rather than necessarily a single task and a mixture.

Four separate cost tariffs

So far, we have compared payoffs without costs. We now ask a different question: does a construction retain its advantage when a particular resource has a price?

We separately examine four assumptions about which resource carries a cost:

Resource pricedNotationWhat the quantity measures
Sensory accessCF=∣F∣−1C_F = \lvert F\rvert- 1The number of distinguishable inputs beyond one
Provisioned capacityCK=K−1C_K = K - 1The number of provisioned categories beyond one, even if not all are used
Representation categoriesCR=∣R∣−1C_R = \lvert R\rvert- 1The number of categories actually used beyond one
Retained world informationCI=I(W;R)C_I = I(W; R)The amount of retained world information in bits

Table 1.

The first three quantities are measured relative to a one-category baseline. These are hypothetical resource measures, not measurements of energy or metabolism. Keeping the cost explicit follows the information–cost distinction relevant to lossy compression [7].

We specify the resource amount and its price separately. Let xx denote a construction or module, ee an objective, and cc the chosen tariff. The quantity Cc(x)C_c(x) gives the resource amount being counted. Its hypothetical price per unit is λ≥0\lambda\ge0.

The product λCc(x)\lambda C_c(x) is the cost expressed in payoff units. Subtracting it from payoff Ve(x)V_e(x) gives net performance:

Jλ(x,e,c)=Ve(x)−λCc(x).J_\lambda(x, e, c) = V_e(x) - \lambda C_c(x).

A high payoff therefore does not guarantee high net performance: costs may outweigh the advantage.

For a pair, let ΔC\Delta C denote the difference in resource amounts and ΔJ\Delta J the difference in net performance. Both differences use the same ”left minus right” order as ΔV\Delta V. Then

ΔJ=ΔV−λΔC.\Delta J = \Delta V - \lambda\Delta C.

There are five cases:

  1. Equal payoffs and resource amounts. Net performances are equal at every price.

  1. Different payoffs with equal resource amounts. The higher-payoff member retains its advantage.

  1. Equal payoffs with different resource amounts. At zero price, net performance is equal; at a positive price, it is higher for the cheaper member.

  1. The higher-payoff member is also more expensive. The leader depends on the resource price.

  1. The higher-payoff member is also cheaper. It retains its advantage at every nonnegative price.

For ΔC≠0\Delta C \ne0, setting ΔJ=0\Delta J = 0 gives

λ∗=ΔVΔC.\lambda^{*} = \frac{\Delta V}{\Delta C}.

Only the fourth case has a positive threshold at which the leader changes. Below λ∗\lambda^{*}, the member with the higher original payoff retains the advantage. At the threshold, net performances are equal; above it, the cheaper member has the advantage.

We retain all five cases. Each tariff is evaluated separately: the four costs are not added together into a single penalty.

Computation and validation of results

The finite sets under study, scoring objectives, and cost rules were fixed in internal configurations before their respective calculations.

Later, during internal manuscript review, we added the representation-once and nested-refinement controls. We also separately counted pairs whose payoffs differ under at least one objective. Worked examples were selected by a fixed ordering, independently of effect size. Algebraically equivalent threshold expressions were standardized. These additional analyses use the same underlying finite sets; Supplementary § 11 describes their chronology and purpose.

Calculations used Python 3.12 and exact rational arithmetic. Information values were stored as expressions q0+q3log⁡23q_{0} + q_{3} \log_{2} 3, with coefficients q0q_{0} and q3q_{3} represented as exact fractions. This representation allows equality and order to be determined without rounding.

Tables and figures are based on computed results saved before the publication summaries were prepared. Summary measures were derived from these records using unambiguous counting rules.

Validation used separate programs with their own implementations of the calculations. They reconstructed the full results: all 15 representations and related policy and transfer sets; all 44,775 modular-analysis rows; and all 916,265 objective- and cost-analysis rows. The checks included threshold signs and all alternatives attaining the same maximum payoff.

Results matched in every checked row. Supplementary § 29 provides details of the validation.

Supplementary tables are numbered in their order of appearance in the Supplement. Tables S1–S2 document provenance and analysis chronology; S3 gives the worked information example; S4–S7 give output counts and their derivation; S8–S12 give comparison denominators and censuses; S13–S15 cover reversals and worked examples; S16–S17 summarize the separate cost tariffs. Section and table numbers are distinct.

Results

Accuracy depends on more than the number of categories

Consider the task of identifying the first, left-hand bit of the world state.

Representation 0011 groups the states into two categories: {00,01}\{00, 01\} and {10,11}\{10, 11\}. In the first category, the left-hand bit is always 0; in the second, it is always 1. Assigning the corresponding answer to each category is therefore enough to solve the task perfectly.

Representation 0102 has more categories — three: {00,10}\{00, 10\}, {01}\{01\}, and {11}\{11\}. However, its first category groups states with different left-hand bits. A single answer cannot be correct for both. Even the best policy answers correctly in only three of the four states.

Thus, two categories allow a payoff of 1, while three categories with a different grouping allow only 3/4. The reason is visible in the groupings themselves: they retain different distinctions between world states.

RepresentationCategories KRK_RWorld information I(W;R)I(W;R), bitsTask information I(Y;R)I(Y;R), bitsInformation in the action I(Y;Aout)I(Y;A_{\mathrm{out}}), bitsBest task payoff
001121.0001.0001.0001
010231.5000.5000.3113/4

Table 3.

The information measures show the same mismatch. Representation 0102 retains more information about the world overall, but less about the correct answer than 0011. In an optimal action for 0102, information about the correct answer amounts to approximately 0.311 bits. This value is rounded; exact expressions and both optimal policies for 0102 are provided in Supplementary Table S3.

Figure 2 shows that each of the three tasks can be solved perfectly with two categories if the grouping matches the task. The complete census of comparisons across all 15 representations shows how common the different cases are within this finite model. Here, one comparison is one pair of representations evaluated on one of the three single tasks.

  • In 12 of 207 comparisons with different category counts, the representation with fewer categories solves the task perfectly, while the one with more categories makes errors.

  • In 66 of 108 comparisons with equal category counts, the best attainable payoffs differ.

  • In 57 of 243 comparisons with different amounts of world information, the representation retaining more world information contains the same amount of information about the correct task answer, or less.

The example in the table is the first of the 12 cases above in the fixed ordering of grouping codes. Tasks are considered in the order: first bit, second bit, XOR.

Finally, having the required information and choosing an appropriate action are distinct conditions for solving a task successfully. We examined how a policy optimal for the source task performs on another task without adjustment.

Even for a representation that distinguishes all four world states, this unchanged policy falls short of the best target-task rule in all six directed transfers: from each of the three tasks to the other two. The representation still contains all the necessary information, but the policy answers the source-task question. Solving the new task perfectly requires only choosing the appropriate action rule.

Refining a representation does not lower its payoff ceiling

In the preceding example, representations 0011 and 0102 group the world states differently. Each preserves distinctions that the other loses. Such representations are called non-nested.

Now consider a different change: preserve all existing distinctions and add new ones. We can do this by splitting existing categories into smaller ones, without merging states from different original categories. The resulting representation is finer. If at least one category is split, the refinement is strict.

For example, take representation 0011 and split category {10, 11} into {10} and {11}, leaving category {00, 01} unchanged. For the left-bit task, both new categories can be assigned the same answer, 1. An additional distinction is available, but the policy need not use it.

In the same way, a finer representation can reproduce the behaviour of any policy available to the coarser one: simply assign the original answer to every part of a split category. When the best policy is chosen and cost is not included, the finer representation’s payoff ceiling therefore cannot be lower. This argument connects to the classical comparison-of-experiments framework, which compares information sources through the decisions they make available [3].

The complete census of strict-refinement pairs across the three single tasks contains 135 comparisons. The finer representation has a higher payoff ceiling in 90 cases and an equal ceiling in 45. There are no cases in which the coarser representation has a higher ceiling (Supplementary Table S12).

Thus, simply removing distinctions from a given representation does not increase its best attainable payoff. Strict lower-information/higher-ceiling cases compare non-nested representations that preserve different distinctions between world states.

This conclusion concerns payoff V∗V^{*} before costs are subtracted. If the finer representation is more expensive under the chosen cost tariff, its payoff can be at least as high while its net performance JλJ_{\lambda} is lower after cost is subtracted.

How coding rules affect the census

Different encodings can specify identical actions. This happens when every answer in the decision code DD is complemented and the action code α\alpha is switched at the same time: the two changes cancel each other out. For example, an answer of 0 first becomes 1, but the switched action code turns it back into 0. The final action remains unchanged for every world state.

The 1,424 complete coded constructions therefore implement 712 distinct combinations of module and action rule across all world states. Each such combination has exactly two encodings.

Each of the 181 modules is evaluated under seven objectives, giving 181×7=1,267181 \times7 = 1{,}267 combinations. In each, we retain all encodings that attain the highest payoff. For each combination, their number is 2, 4, 8, or 16. It is always even because every encoding has a partner that specifies the same actions. The absence of a unique optimal encoding therefore does not necessarily mean that there are several distinct optimal ways to act.

A different feature of the coding explains the equal counts of the two types of advantage under representation replacement. The complete census contains 9,912 rows, each evaluating one pair of constructions under one objective:

  • in 5,880 rows, the amount of retained world information is equal;

  • in 792 rows, the lower-information member of the pair has a higher payoff;

  • in another 792 rows, the higher-information member has a higher payoff;

  • in 2,448 rows, information differs but payoffs are equal.

As explained in Section 2.5, both members of each pair have the same number of categories. Their membership and sizes, and hence retained information, can differ.

Why does each of the two opposing advantages occur exactly 792 times? Take any comparison pair and switch only the action code α\alpha in both constructions, leaving all other components unchanged. Every action now becomes its opposite. The representations and the amount of retained information remain unchanged, but correct answers become incorrect and incorrect answers become correct.

Each construction’s payoff therefore changes from VV to 1−V1-V. This holds for both a single task and a task mixture, which averages the task payoffs. The payoff difference changes sign:

(1−V1)−(1−V2)=−(V1−V2).(1-V_1)-(1-V_2)=-(V_1-V_2).

In other words, if the lower-information member previously had the higher payoff, the higher-information member has the higher payoff after the switch, and vice versa.

The full enumeration includes both action codes. Every comparison with unequal information and unequal payoffs therefore has a partner with the opposite advantage. There are 1,584 such comparisons: 792 in each direction. The equality also holds when each objective is counted separately.

The equal counts thus follow from the model’s rules: binary actions, correct/incorrect scoring, and the inclusion of both action codes in the enumeration. With a different action set, different coding rules, or a different scoring system, this equality is no longer guaranteed.

The shares depend on whether we count modules or representations

We now compare what modules can achieve after choosing the best action rule for each. Under each objective, we optimize their decision and action codes separately. In the preceding representation-replacement comparison, these codes remained fixed.

For this census, it matters whether we count modules or representations. Two modules can group world states identically while differing in their sensory filter or provisioned capacity. For example, one module may have spare capacity for unused categories while another has none. These are different modules, even though their retained representation is the same.

In the set of modules studied, a representation appears in between 1 and 60 modules. Some representations therefore enter the all-module comparison repeatedly. To determine how these repetitions affect the result, we compare two counting rules:

  • All module pairs. The 181 modules give 16,290 pairs. Each is evaluated under seven objectives, yielding 114,030 rows.

  • Representation pairs without repeated implementations. We retain one instance of each of the 15 representations. This gives 105 pairs and 105 ×\times 7 = 735 rows.

The second census uses the same saved results. Only the counting rule changes: differences in sensory filters and provisioned capacity no longer produce repeated comparisons of the same representations.

The table summarizes the two complete censuses across all seven objectives. Each data row belongs to one of four classes (Supplementary Tables S10 and S11).

Information and payoff-ceiling relationshipAll module pairs: rowsDistinct representation pairs: rows
Equal information, regardless of payoff31,122168
Lower information, higher payoff ceiling7,34460
Higher information, higher payoff ceiling51,891296
Unequal information, equal payoff ceilings23,673211
Total114,030735

Table 1 (PDF p. 16).

Both censuses contain cases in which the lower-information member has the higher payoff ceiling and cases in which the higher-information member has the advantage. Under both counting rules, the latter cases outnumber the former.

The shares of these cases nevertheless differ. The first census characterizes all admissible modules, including different implementations of the same representation. The second characterizes the ways of grouping world states, counting each once. Any statement about which advantage occurs more often must therefore specify the counting rule used to obtain the shares.

Figure 4 compares three censuses: representation replacements, all module pairs, and representation pairs without repetitions. Results are summed over all seven objectives. Bar length shows the class’s share of all rows in that census, and the adjacent label gives the exact row count.

Figure 4. Relationships between retained world information and payoff under three counting rules. For representation replacements, equal counts of opposing advantages follow from coding symmetry. The all-module comparison counts different implementations of a representation; the representation comparison without repetitions counts each representation once. All four classes are shown separately for each counting rule, including equal information and equal payoffs at unequal information.

The advantage depends on which tasks are evaluated

The same construction can outperform another on one task and fall behind when several tasks are evaluated together. The example in Supplementary Table S15 illustrates this. It was selected by a fixed ordering, not by effect size.

ComparisonObjectiveLeft payoffRight payoffLeft minus right
Fixed construction pairSecond-coordinate task1/43/4-1/2
Same construction pairEqual first-coordinate/XOR mixture1/21/41/4
Module pairFirst-coordinate task; information tariff1/23/4-1/4

Table S15. Worked objective-reversal and information-cost examples.

The two constructions in this example differ only in their representation. The left construction uses grouping 0001: it combines states 00, 01, 10 and distinguishes 11 separately. The right construction uses grouping 0010: it combines 00, 01, 11 and distinguishes 10 separately. All other components are identical.1 The labels “left” and “right” remain attached to these constructions under every objective, as described in § 2.7.

Tasks evaluatedLeft construction’s payoffRight construction’s payoff
Right-bit task1/43/4
Equal mixture of the left-bit and XOR tasks1/21/4

Table 4.

The right construction has the higher payoff on the single task, whereas the left has the higher payoff on the mixture. The left-minus-right payoff difference is −1/2-1/2 and +1/4+1/4, respectively. Both constructions and their action rules remain unchanged; only the set of tasks used to evaluate performance differs.

We count an advantage reversal when at least one single task and at least one task mixture favour different members of the pair. The payoffs must differ in both cases: equal payoffs alone do not establish a reversal.

The complete census of 9,568 pairs differing in one component identifies 2,116 pairs with an advantage reversal.

Not all admissible component replacements affect payoff, however. Many pairs have equal payoffs under all seven objectives. Retaining only pairs whose payoffs differ under at least one objective leaves 3,592 pairs. These are called payoff-active in the supplementary materials.

The same 2,116 pairs reverse within this set of 3,592. The two denominators therefore answer different questions:

  • 2,116 of 9,568: how many pairs reverse among all admissible one-component replacements?

  • 2,116 of 3,592: how many reverse among replacements whose payoffs differ under at least one objective?

Supplementary Table S14 in § 24 gives the detailed census of active and remaining pairs for each component type.

Component familyIdentitiesActive in ≥1\ge1 objectiveInert in all 7Reversals
sensory filter3,26403,2640
provisioned capacity1,32001,3200
representation1,416748668676
decision code2,8562,4943621,272
action code712350362168
representation/decision/action family subtotal4,9843,5921,3922,116
actually payoff-active identities3,5923,59202,116

Table S14. Payoff-active and inactive comparisons by component.

For modules, we choose the best action rule separately under each objective. Among 16,290 module pairs, 3,672 pairs reverse their advantage between a single task and a mixture.

An additional, broader check considers any two of the seven objectives: does one favour one module while another favours the other? The comparison can involve two single tasks or two mixtures as well. This condition holds for 4,395 module pairs, including the 3,672 pairs identified above.

The reversal threshold depends on the chosen cost tariff

We now keep the task fixed and account for costs. A module may solve the task more accurately but require more of the resource being priced. Its payoff advantage must then be weighed against the additional cost.

To do this, we subtract the cost of the resource used from payoff VV:

Jλ=V−λC.J_{\lambda} = V - \lambda C.

Here, CC is the resource amount under the chosen tariff, and λ\lambda is the hypothetical price per unit of that resource, expressed in payoff units. The quantity JλJ_{\lambda} is net performance after cost is subtracted.

If the higher-payoff module is also more expensive, it retains its advantage at a low price. At a sufficiently high price, the extra payoff no longer compensates for the additional cost, and the cheaper module becomes the better option. The boundary is an exact positive threshold,

λ∗=ΔVΔC>0.\lambda^* = \frac{\Delta V}{\Delta C} > 0.

At this price, net performances are equal. Below the threshold, the higher-payoff module has the advantage; above it, the lower-cost module does.

For each of the four tariffs, we separately examine the complete finite set of 114,030 module-pair/objective rows. The table reports how many rows have such a positive threshold.

Resource pricedResource amount CCRows with a positive threshold
Sensory distinctionsCF=∣F∣−1C_F = \lvert F\rvert- 131,415
Provisioned capacityCK=K−1C_K = K - 130,414
Representation categoriesCR=∣R∣−1C_R = \lvert R\rvert- 145,951
Retained world informationCI=I(W;R)C_I = I(W;R)51,891

Table 5.

The four tariffs are separate alternatives; their costs are not added together. Both the counts and the threshold values differ. Exact ranges and summary statistics are reported in Supplementary Table S17 in § 26.

Consider an example for the left-bit task. This is the first pair with a positive threshold under the information tariff in the fixed module enumeration order.

The left module uses representation 00000000, grouping all four world states into one category. It retains 0 bits of information and attains payoff 1/21/2.

The right module uses representation 00010001, distinguishing state 1111 from the other three. It attains a higher payoff, 3/43/4, and retains

I(W;R)=2−34log⁡23I(W;R) = 2 - \frac{3}{4}\log_2 3

bits of information.2

The right module’s payoff advantage is 1/41/4. The cost of its additional information exactly offsets this advantage at a price of

λ∗=1/42−34log⁡23=18−3log⁡23≈0.3082\lambda^* = \frac{1/4}{2 - \frac{3}{4}\log_2 3} = \frac{1}{8 - 3\log_2 3} \approx0.3082

payoff units per bit. Above this price, the left module has higher net performance. Its payoff still equals 1/21/2: it has not become more accurate at the task, but it is the better option after costs are included.

The count of 51,891 under the information tariff matches the number of higher-information/higher-ceiling cases in Section 3.4. These are the same comparison rows. With CI=I(W;R)C_I = I(W;R), more information automatically means more of the resource being priced. Each such case therefore has a positive threshold at which the advantage reverses. The equal counts follow from the tariff’s definition and are not a separate, independent confirmation of the result.

Figure 5 brings together the two ways of changing the advantage. In part A, the set of evaluated tasks differs. In part B, the task remains fixed, but the price of the chosen resource is included. The first comparison concerns payoffs; the second concerns net performance after cost.

Figure 5. Part A shows the number of pairs whose advantage reverses between a single task and a mixture. For one-component replacements, counts are reported among all admissible pairs and among payoff-active pairs; for modules, the broader check across any two objectives is also shown. Part B shows the number of comparisons with a positive reversal threshold under each of the four separate tariffs.

Discussion

Task performance depends on which distinctions a representation preserves

The example in § 3.1 shows why more retained world information does not always allow greater task accuracy. A representation can preserve many distinctions while merging precisely the states that require different answers. Another representation may retain less information overall but distinguish states that matter for the task, thereby providing a higher payoff ceiling.

A different situation arises when a representation preserves all previous distinctions and only adds new ones. A policy can leave the additional distinctions unused and reproduce the previous actions. With no cost and the best policy chosen, the payoff ceiling therefore does not decrease. The lower-information representation’s advantage in the first example comes from retaining different distinctions, not simply from deleting some information from a given representation.

Accounting for costs introduces another basis for a lower-information advantage. If the lower-information module is cheaper under the chosen tariff, the savings may outweigh its task-performance deficit. Its payoff VV can then be lower while its net performance after cost, JλJ_{\lambda}, is higher. The advantage here comes from lower costs, not greater accuracy.

How often the different cases occur also depends on which comparisons are included in the census. Under representation replacement, lower-information/higher-payoff and higher-information/higher-payoff cases occur in equal numbers because of coding symmetry. In module comparisons, the shares depend on how many different modules implement each representation. Reported shares must therefore specify which constructions were compared and what counted as a distinct member of a comparison pair.

The set of evaluated tasks matters as well. The same constructions with unchanged action rules can have different advantages depending on whether a single task or a mixture of tasks is evaluated. The usefulness of retained distinctions depends not only on the representation but also on the tasks to be solved.

The separation of information quantity from decision value is established in earlier work [13, 11]. In particular, our nested-refinement result is a deterministic special case of the Blackwell order, not a new theorem. This work also builds on studies involving multiple objectives or environmental change [6, 1, 2, 8]. The atlas brings these questions together in one finite model with a complete enumeration of admissible constructions. This makes it possible to compare lower- and higher-information advantages, retain cases of equal payoff, and explain which results follow from the model and counting rules.

How the results relate to Hoffman’s ideas

Hoffman’s ideas led us to ask whether better task performance necessarily requires more information about the world. Our atlas contains cases in which a representation retains less information but supports a higher payoff. In this sense, the results are consistent with successful action based on selectively retained information.

The atlas also contains cases in which more informative representations have the advantage. For nested representations, a more definite result holds: preserving all previous distinctions and adding new ones cannot lower the best attainable payoff before costs are included. The results are therefore compatible with both the loss of some information and the preservation of useful distinctions. The answer depends on which distinctions are retained, which tasks are evaluated, and which costs are included.

To relate these findings to the fitness-beats-truth argument, it is important to distinguish the quantities involved.

In this study, payoff is a score for solving a task. It does not measure survival or reproduction. To determine how an observed advantage affects the spread of constructions in a population, we would need to specify a link between payoff and reproduction and a model of evolutionary dynamics.

Retained world information I(W;R)I(W; R) also has a specific meaning: it measures how much knowing the representation reduces uncertainty about the state of the modelled world. The number of bits alone does not determine whether a description of the world is true or identify which relations it preserves. Two representations with the same amount of information can distinguish different states and help solve different tasks.

Assessing structure preservation requires additional criteria. For example, one could separately test whether a representation reflects causal relations—what affects what—or spatial relations—what is adjacent to what. This kind of question, considered by [10], differs from our comparison of information amount and payoff.

Finally, the fitness-beats-truth theorem is formulated for particular strategy classes and mathematical conditions [9]. Our atlas uses a different finite model, so its results should not be treated as a direct test of that theorem. It supplies a worked finite comparison in which the effects of action rules, counting, and costs can be inspected before introducing a population model.

What the finite atlas contributes

The atlas is intended as a reference for checking comparisons, rather than as a new theory of compression or a new optimization method. For example, a reader can check whether a reported lower-information advantage persists after the action rule is optimized, or whether a census changes when repeated realizations are counted only once. The saved results make these checks possible within the stated model; their usefulness as a benchmark for other methods remains to be evaluated.

This focus differs from mapping the best compression-relevance trade-offs [15, 14]. Our pairwise cost thresholds include comparisons between constructions that need not be globally optimal. Beating one competitor at a given price does not establish superiority over every admissible alternative. The four tariffs expose different modelling choices, not newly discovered mechanisms of resource use.

The model’s small size makes it possible to examine all admissible constructions and all comparisons specified by its rules. This reveals the cases in which a less informative representation has the advantage, those in which a more informative one does, and those in which the results are equal.

Exact arithmetic reliably distinguishes equality from differences in the results. Equality in this census is established from exact values: a match between rounded numbers is not enough. This matters both for comparing payoffs and for identifying the price at which net performances become equal.

Complete enumeration connects the numerical results to the model’s rules. Supplementary § 27 gives assumptions and short proofs for four properties: the ability to ignore additional distinctions, action-inversion symmetry, the effect of repeated encodings on pair counts, and the emergence of a threshold when information is priced. These properties help explain the relationships found in the atlas.

The effect of repeated encodings deserves a separate explanation. Even multiplying the number of encodings of every model phenotype by the same factor can change the comparison shares. Here, a phenotype is a module together with its actual actions in every world state. If the census includes pairs of equivalent encodings of one phenotype, it counts two kinds of pairs: within a phenotype and between different phenotypes. Duplication increases their counts in different proportions. The same increase in encoding multiplicity therefore need not preserve the shares of the results.

These mathematical explanations help account for the observed relationships. The resulting shares describe the complete census of the specified finite set of comparisons under the chosen counting rules.

Treating costs separately also matters. The number of distinguishable sensory inputs, provisioned capacity, the number of used categories, and retained information describe different resources. The four tariffs allow the consequences of each cost assumption to be examined separately.

Applying these tariffs to an organism or device would require identifying the physical process responsible for the costs, their measurement units, and their relationship to the chosen model quantity. The hypothetical price could then be linked to measurements of a specific system.

Related questions arise in learning visual representations. For example, Franca trains nested portions of a feature vector so that its truncated versions also remain useful [16]. Our atlas offers a simple formal model in which retained distinctions, available decision rules, and representation cost can be examined separately. Extending these arguments to learned systems requires accounting for the training procedure and the limitations of the decision rule used.

From comparing constructions to population evolution

The atlas allows us to compare how well different constructions perform tasks. The next step for the five-component constructions in this paper is to test how these differences affect the spread of variants in a reproducing population.

This requires specifying how payoff relates to reproductive fitness—expected reproductive success. Population size and the rules of inheritance, mutation, and reproduction must also be defined. It would then be possible to estimate the probability of fixation, when one variant occupies the entire population, or the shares of different variants over the long term.

For example, fitness could be assigned positive values such that higher payoff always corresponds to higher fitness. This would preserve pairwise order: the construction with higher payoff would also have higher fitness. That relationship alone, however, would not determine its fixation probability or long-run share of the population.

The atlas results help identify further questions. Would different encodings with identical behaviour be inherited as distinct variants? Which mutations would connect them to other constructions? How should genetic drift—random changes in variant frequencies during reproduction—be modelled? If tasks change over time, the sequence of those changes must also be specified: a task mixture with fixed weights does not describe such a sequence.

The exact comparisons in the atlas can serve as benchmarks for this model. The new test would show when an advantage in task performance leads to the spread of a construction and when population dynamics change the outcome.

Limitations

Which model was completely enumerated. Complete enumeration applies to a model with four equiprobable world states and three binary tasks. The tasks are balanced: each of the two answers is correct in two of the four states. Representations and policies are deterministic—the same input always produces the same output.

The results do not cover larger worlds, unequal state probabilities, random noise in representations, learning, or interaction with the environment through a body and its movements. Task mixtures have fixed weights: they do not describe environmental change over time, memory, or adaptation.

What the quantities measure. Retained world information measures how much a representation reduces uncertainty about the world state. Payoff is average answer accuracy, and net performance is payoff minus the specified cost.

These quantities do not provide a general assessment of a representation’s veridicality. Nor does the model establish how payoff relates to reproductive success or how cost relates to actual energy use, memory use, or computation time.

How the rules affect comparisons and counting. When one component is replaced, the other four remain fixed. The result therefore characterizes that replacement in a particular construction; it does not establish a universal effect of the component.

The enumeration rule specifies which constructions are compared, but not how often mutation could turn one into the other or how many mutational steps this would require.

In the module census, a representation may be counted repeatedly; in the representation-once census, it is counted only once. The resulting shares depend on that choice. They describe a complete census of the specified comparisons, not the prevalence of such constructions among natural agents.

Which evolutionary conclusions remain beyond the model. A pairwise payoff advantage does not show whether a rare variant will spread, occupy the entire population, disappear, or coexist with other variants. Such conclusions require a population model of the kind outlined in § 4.4.

This work continues a research question from the fitness–truth literature, but does not reproduce published evolutionary simulations.

Reproducibility and AI use

An accompanying package supports checking and repeating the calculations. It contains the manuscript, supplement, source tables, vector figures, calculation code, separate programs for checking the results, and selected tests. The package also includes model specifications, fixed calculation settings, and records of the exact software dependency versions.

The archive preserves the original computational results and the summaries derived from them. Null results, equal payoffs or net performances, and order reversals are included alongside cases where one construction has a strict advantage over another.

The package includes a manifest listing the files, their locations, sizes, and checksums. A checksum serves as a digital fingerprint of a file’s contents and allows changes to be detected. Supplementary § 29–§ 32 describe the checks, the links between claims and results, and the workflow.

The package was validated in the Python (CPython) environment on Windows described in the manifest. Producing byte-identical files on other operating systems has not yet been independently verified. The texts of primary research papers are not included in the package.

OpenAI Codex was used for source organization, code development and testing, manuscript drafting and editing, and internal consistency checks. The first author reviewed the scientific decisions and interpretations and is responsible for the manuscript’s content. The use of AI in internal checks is described in more detail in the supplement.

Conclusion

To assess how well a representation supports a task, knowing only how much world information it retains is not enough. The particular distinctions it preserves matter. In the four-state model studied here, a representation retaining less information can achieve higher accuracy if it better preserves the distinctions needed for the task. Cases favouring more information and cases of equal accuracy also occur.

Simply discarding distinctions, however, does not improve the best attainable result when cost is excluded. If one representation preserves all the distinctions of another and adds new ones, an optimal policy can ignore the additions and repeat the earlier actions. Additional distinctions therefore do not reduce the payoff ceiling.

Advantage also depends on the task and costs. Changing the task mixture can change which of two unchanged constructions receives the higher payoff. Once cost is included, the net-performance advantage can shift to a cheaper construction even though its answer accuracy remains unchanged. The shares of such cases depend on which pairs are compared and whether repeated implementations of one representation are counted separately.

The relationship between information and payoff must therefore be considered together with the retained distinctions, task, available decision rules, cost, and counting method. These general distinctions are established in earlier theory. Our contribution is an exact, reproducible account of their consequences in one fully enumerated model, including equal outcomes, coding symmetries, and alternative counting rules. The saved comparisons can serve as reference cases for subsequent modelling; they do not establish a new general advantage of information loss or an outcome of population evolution.

Author contributions

Aleksandr M. Shtraukh-Adamsky led the study and made the principal contributions to conceptualization, model development, computational experiment design, computational implementation, validation and analysis of results, and manuscript drafting and editing. The use of AI tools is described in § 6.

Vladimir M. Adamsky, Irina G. Voloshina, and Artemiy D. Spitsyn contributed to the discussion and development of the research concept, models, and computational experiment designs.

All authors have read and approved the manuscript.

Short glossary

Representation and category. A representation groups world states. States in the same category receive the same internal label, so the representation does not distinguish them.

Retained world information. I(W;R)I(W;R) measures how much knowing the category reduces uncertainty about the world state on average. It is measured in bits.

Bit of information. A unit of information. Learning which of two equiprobable possibilities has occurred provides one bit. Average information can be fractional; the introduction explains the value of 1.5 bits with a worked example. Task and objective. A task specifies which answer is correct. An objective evaluates accuracy on one task or average accuracy across several tasks with specified weights.

Policy. A rule assigning an action to each representation category.

Payoff and payoff ceiling. VV is average accuracy on a specified task or task mixture. For example, V=3/4V = 3/4 corresponds to 75% correct answers. Equal payoffs mean equal accuracy under the same evaluation conditions. V∗V^* is the highest accuracy attainable by choosing the best allowed policy for the given representation or module.

Module. A combination of three components: a sensory filter defining the distinctions available at the input, provisioned capacity, and a representation. Decision and action codes have not yet been selected.

Capacity and category count. KK is the maximum allowed number of representation categories. KRK_R is the number actually used. For example, a representation with capacity K=4K = 4 may use only two categories.

Nested representations. Two representations are nested if one preserves every distinction made by the other. The finer representation may further split the coarser one’s categories, but does not merge states that the coarser one already distinguishes.

Comparison pair and row. A pair consists of two constructions or two modules being compared. A comparison row evaluates that pair under one objective. One pair evaluated under seven objectives therefore produces seven rows.

Complete census. A count of every case specified by the chosen rules, such as all admissible constructions or all pairs being compared.

Resource, price, and net performance. CcC_c is the resource quantity under the chosen tariff, and λ\lambda is its unit price. The cost λCc\lambda C_c is subtracted from payoff to obtain net performance:

Jλ=V−λCc.J_\lambda= V - \lambda C_c.

Crossing threshold. The price λ∗\lambda^* at which the two compared endpoints have equal net performance. If the more accurate endpoint is also more expensive, it has the advantage below this threshold; above it, the cheaper endpoint does.

An accompanying reader guide provides a longer glossary and answers to common questions.

Supplementary Material

When Does Less Information Pay? An Exact Finite Atlas of Representation, Payoff, and Cost

Status: Supplement to preprint version 1.0. License: CC BY 4.0.

Task: TASK-PUB-005

Claim: CLM-PUB-001

Epistemic status: E

Scope and provenance

The paper synthesizes only accepted Stage 7.2A–7.2C results:

Table S1 documents the scope and fixed inputs of the three analyses.

StageDecisionAssumptionClaimExact config SHA-256
7.2ARDR-0028ASM-INT-0002CLM-INT-BOTTLENECK-001a1feca94512b19023e518a6000de7901cebb17e7 2d387873cc0d397ad94e6622
7.2BRDR-0029ASM-INT-0003CLM-INT-GENOTYPE-0010f2fe87d1347fcfd648bcdceb39395df91617ef2 ad0926227ca86f6f5dd9298e
7.2CRDR-0030ASM-INT-0004CLM-INT-SELECTION-BRIDGE-0015d3f2e25a80b60d9e89d4abbfaed5c4c295f44eb 9f55f2bb00a3558cef40c8ef

Table S1. Scope and provenance of the three exact analyses.

The paper is a project-defined exact finite extension. It is not a direct reproduction of Mark et al. [5], Prakash et al. [9], or any source genetic algorithm. The Hoffman book is motivation only. The renewed source audit found the public Berke et al. [2] OSF code/data package and therefore limits the recorded reproducibility gap to the older materials examined.

The study was not publicly preregistered. The primary finite sets of constructions and comparisons, the objectives, and the tariffs were fixed in the exact stage configurations before their executions. The following saved-evidence analyses were added later during manuscript review:

Table S2 records the purpose of the later saved-evidence analyses.

Table S2. Analyses added during manuscript review.

Later analysisReason addedScientific effect
representation-once quotientexpose multiplicity weighting in the 181-module censusalternative counting measure over the same realized representations; no new model rows
strict nested-refinement censustest whether lower-information witnesses arise by merely deleting distinctionsnarrows the mechanism claim; no new model rows
active/inert denominator decompositiondistinguish all categorical replacements from those that ever change payoffclarifies denominators only
information-decomposition and order examplesmake the information and reversal mechanisms readableselected mechanically, never by effect magnitude
exact threshold canonicalizationcollapse algebraically equivalent symbolic quotientschanges the number of distinct spellings/classes, not any saved threshold row

Table 6.

The dated search procedure behind the limited reproducibility-gap statement is packaged as provenance/PAPER-001-hoffman-motivation-reproducibility-gap-audit.md.

Exact finite model

World states are 00, 01, 10, 11, each with probability 1/41/4. Actions are 0, 1. The three task label vectors are:

bit_0 = (0,0,1,1)
bit_1 = (0,1,0,1)
xor   = (0,1,1,0)

In reader-facing prose, bit_0 is the first-coordinate task (the left bit in the declared world ordering) and bit_1 is the second-coordinate task (the right bit). These three tasks exhaust balanced binary labelings of four states up to output complement: choose⁡(4,2)/2=3\operatorname{choose}(4,2)/2=3.

This completeness statement concerns individual task partitions. Multi-task objectives use the three answer labelings listed above without independent inversions. Because the same action rule serves every task in a mixture, inverting one task’s answers can change the payoff of a fixed representation.

Payoff is one for a correct action and zero otherwise. A deterministic representation is a label-free partition of the four states encoded as a restricted-growth string. The 15 canonical identities are available in source_tables/fig2_representation_atlas.csv.

For deterministic R=r(W)R=r(W), retained world information is I(W;R)=H(R)I(W;R)=H(R) with base-2 entropy. Exact values use q0+q3log⁡2(3)q_0+q_3\log_2(3), where q0q_0 and q3q_3 are rational coefficients stored in each symbolic information certificate. Display decimals are non-normative.

Exact values for the introductory example

The following saved example concerns the first-coordinate task. Policy bits are listed in category-label order. Both optimal policies for 0102 have the information-in-action value shown. The example was selected by the fixed ordering described in main-text § 3.1.

Table S3 gives the exact values and optimal policies for the introductory example.

QuantityRepresentation 0011Representation 0102
Used categories23
Category labels in world order[0,0,1,1][0,1,0,2]
Retained world information, bits13/2
Task-relevant information, bits11/2
All optimal policies[[0,1]][[0,0,1],[1,0,1]]
Information in an optimal action, bits1(3/2)+(−3/4)∗log⁡2(3)(3/2)+(-3/4)*\log_2(3)
Best task payoff13/4

Table S3. Exact information decomposition of the introductory example.

The machine-readable values and selection rule are in source_tables/table_s5_information_decomposition_witness.csv.

Complete output counts

Stage 7.2A

Table S4 lists the saved output counts for the representation atlas.

TableRows
representations15
representation-policy-task282
representation-task ceilings45
frozen transfer90
active-category thresholds315
information-rate thresholds315
failures0

Table S4. Output counts for the representation atlas.

Counterexample counts: 12 smaller-K_R perfect/larger-K_R imperfect representation-pair/task comparisons out of 207 different-K_R comparisons; 66 same-K_R unequal-ceiling comparisons out of 108; 57 of 243 unequal-I(W;R) representation-pair/task comparisons retain more world information without greater task-relevant information; six full-state frozen-policy target losses.

Stage 7.2B

Table S5 lists the saved output counts for the modular-construction atlas.

TableRows
compatible modules181
genotypes1,424
genotype-task4,272
module frozen-transfer envelopes1,086
contextual one-locus contrasts28,704
epistasis rectangles9,108
failures0

Table S5. Output counts for the modular-construction atlas.

Derivation of the central library counts

For construction g=(F,K,R,D,α)g=(F,K,R,D,\alpha), the realized action is Ag(w)=α(D(R(w)))A_g(w)=\alpha(D(R(w))). Compatibility requires RR to coarsen FF and ∣R∣≤K|R|\le K.

Table S6 gives the derivation of the central construction counts.

CountExact derivation
181 compatible modulescapacity subtotals 15+47+59+60=18115+47+59+60=181 for K=1,2,3,4K=1,2,3,4
1,424 coded constructionsthe 181 modules have realized-∣R∣|R| distribution 60, 96, 24, 1; therefore 60×22+96×23+24×24+1×25=1,42460 \times2^2+96 \times2^3+24 \times2^4+1 \times2^5=1,424, where the extra factor 2 is the action code
9,568 one-coordinate identitiesunordered alternative pairs within every fixed four-component background: 3,264+1,320+1,416+2,856+712=9,5683,264+1,320+1,416+2,856+712=9,568 for F,K,R,D,αF,K,R,D,\alpha

Table S6. Combinatorial derivation of the central construction counts.

These are combinatorial counts in the declared library. They are not sampled abundances or population sizes.

An epistasis rectangle fixes a background and two binary component substitutions, then records the non-additive contrast ϵ=V11−V10−V01+V00\epsilon=V_{11}-V_{10}-V_{01}+V_{00}. It is a project-defined contextual finite-difference diagnostic, not biological epistasis or a population parameter.

The 1,424 genotypes encode 712 distinct module-by-realized-action phenotypes, each with two neutral decision/action encodings. Complementing the decision code and flipping the action code preserves the realized action map, so this twofold multiplicity is a coding invariant.

Stage 7.2C

Table S7 lists the saved output counts for the objective and cost analyses.

TableRows
genotype-environment profiles9,968
module-environment envelopes1,267
registered one-coordinate/environment comparisons66,976
module pair-environment comparisons114,030
registered one-coordinate threshold rows267,904
module threshold rows456,120
total scientific rows916,265
failures0

Table S7. Output counts for the objective and cost analyses.

Environments

The seven environments are all nonempty equal-weight subsets of the three tasks:

stable_bit_0

stable_bit_1

stable_xor

mix_bit_0_bit_1

mix_bit_0_xor

mix_bit_1_xor

mix_all_tasks

Mixtures are stationary expectations, not temporal switching processes.

Comparison universes and exact witness counts

The same comparison pair can contribute one row per objective. The following inventory gives the denominators used in the article.

Table S8 distinguishes comparison pairs from objective-specific rows.

Comparison setPairsObjectivesComparison rows
All registered one-component replacements9,568766,976
Representation, decision-code, and action-code family subtotal4,984734,888
Actually payoff-active registered replacements3,592725,144

Table S8.

Table S8. Comparison pairs, objectives, and row denominators.

Comparison setPairsObjectivesComparison rows
Registered representation-component replacements1,41679,912
Unrestricted module-ceiling pairs16,2907114,030
Realized representations counted once1057735

Table 7.

Full definitions are in source_tables/table_s6_denominator_map.csv.

Registered representation-component replacements

Across 9,912 rows:

The two representations in each pair have the same number of categories: the fixed decision code has one bit per used category. Unequal retained information in this census comes from different block sizes at a fixed category count, not from comparing a representation with a strict refinement of itself.

Table S9 reports the four classes for representation replacements.

ClassCount
equal world information5,880
lower world information, higher payoff792
higher world information, higher payoff792
unequal world information, payoff tie2,448

Table S9. Information and payoff classes for representation replacements.

Module-ceiling comparisons

Across 114,030 rows:

Table S10 reports the four classes for module-ceiling comparisons.

ClassCount
equal world information31,122
lower world information, higher payoff7,344
higher world information, higher payoff51,891
unequal world information, payoff tie23,673

Table S10. Information and payoff-ceiling classes for module pairs.

These are complete censuses of the specified finite sets of comparisons. They are not probabilities or empirical frequencies.

Representation-once quotient

The 181 modules weight realized representations unequally: a realized map has between 1 and 60 compatible sensory-filter/capacity encodings. A deterministic aggregation of the immutable saved rows therefore counts each realized representation once in each environment. Across 735 comparisons:

Table S11 reports the census with each representation counted once.

ClassCount
equal world information168
lower world information, higher payoff60
higher world information, higher payoff296
unequal world information, payoff tie211

Table S11. Information and payoff-ceiling classes with each representation counted once.

This quotient was added during manuscript review to expose the counting measure, not to replace an unfavourable headline. It is reported beside the original module census and is derived deterministically from the same immutable rows.

Nested-refinement control

For a strict refinement, every state pair merged by the finer representation is also merged by the coarser representation. Across the complete 135 strict-refinement/task comparisons:

Table S12 reports the strict nested-refinement control.

Relation of reoptimized ceilingsCount
finer representation higher90
tie45
coarser representation higher0

Table S12. Strict nested-refinement control.

The zero follows from policy-set inclusion: the finer representation can imitate any policy available to the coarser one. Strict lower-information/higher-ceiling witnesses therefore require non-nested partitions in this atlas.

Reversal observable

For a fixed left/right identity, let delta_value=V_left-V_right. The registered single-task-versus-multi-task diagnostic requires one strictly positive sign in the single-task set and one strictly negative sign in the multi-task set, or the reverse. Zero establishes neither side. The orientation is never changed by environment.

Table S13 reports the single-task versus task-mixture reversal counts.

UniverseUnique identitiesSingle-task-vs-multi-task reversals
registered one-coordinate comparisons9,5682,116
module pairs16,2903,672

Table S13. Single-task versus task-mixture order reversals.

The broader “both signs anywhere among seven environments” module diagnostic is 4,395/16,290 and is not interchangeable with 3,672.

The 9,568 denominator includes two component families that are structurally payoff-inert in the registered composition: 3,264 sensory-filter identities and 1,320 capacity identities. A deterministic decomposition of the immutable rows is:

Table S14 separates payoff-active and inactive comparisons by component.

The registered graph contains every categorical alternative at one component and fixed background. It does not define a transition process, mutation rate, or distance.

Worked examples

The first two rows evaluate the same fixed pair under different objectives. The third row is the separate cost example from main-text § 3.6.

Table S15 provides the saved objective-reversal and cost-threshold examples.

For the construction pair, the common components are F=0012F=0012, K=2K=2, D=01D=01, and the bit-flip action code. The left representation is 00010001; the right is 00100010. For the module pair, the left module is F=0000F=0000, K=1K=1, R=0000R=0000 and the right is F=0001F=0001, K=2K=2, R=0001R=0001. Their information costs are respectively 00 and (2)+(−3/4)∗log⁡2(3)(2)+(-3/4)*\log_2(3). The positive threshold is 1/((8)+(−3)∗log⁡2(3))1/((8)+(-3)*\log_2(3)), approximately 0.3081557267020.308155726702. Exact identities, costs, and the unchanged selection rules are in source_tables/table_s7_worked_order_examples.csv.

Tariffs and exact threshold algebra

Four costs are evaluated independently:

Csensory=∣F∣−1C_{\mathrm{sensory}} = |F|-1
Ccapacity=K−1C_{\mathrm{capacity}} = K-1
Crep_support=∣R∣−1C_{\mathrm{rep\_support}} = |R|-1
Crep_information=I(W;R)C_{\mathrm{rep\_information}} = I(W;R)

For each tariff, VV is zero-cost task payoff, CcC_c is the amount of the named mechanism basis, λ\lambda is its hypothetical unit price, and Jλ=V−λ∗CcJ_\lambda=V-\lambda*C_c is tariff-adjusted net performance. JλJ_\lambda is not called payoff. The exact pairwise boundary is λ∗=δvalue/δcost\lambda^*=\delta_{\mathrm{value}}/\delta_{\mathrm{cost}} when cost differs. Five exhaustive classes are retained: coincident; equal-cost payoff dominance; zero-price tie followed by lower-cost dominance; positive crossing; payoff-and-cost dominance.

Module positive-crossing counts:

Table S16 reports positive crossings and payoff-and-cost dominance separately for each tariff.

TariffPositive crossingsPayoff-and-cost dominance
sensory support31,41510,983
capacity provision30,41416,035
representation support45,9511,404
representation information51,8917,344

Table S16. Positive crossings and payoff-and-cost dominance by tariff.

The tariffs are never pooled and none is interpreted as measured biological cost.

Positive-threshold distributions:

Table S17 summarizes the positive threshold distributions.

TariffDistinct valuesMinimumMedianMaximum
sensory support81/361/41/2
capacity provision81/361/61/2

Table S17. Distributions of positive module-pair thresholds.

TariffDistinct valuesMinimumMedianMaximum
representation support61/241/241/41/41/21/2
representation information17≈0.0833\approx0.0833≈0.3082\approx0.3082≈1.3247\approx1.3247

Table 8.

Exact symbolic information thresholds are in source_tables/table_s4_threshold_distributions.csv. No empirical range for lambda is asserted. Because the information tariff defines C=I(W;R)C=I(W;R), its 51,891 positive crossings are algebraically the same set as the 51,891 higher-information/higher-payoff module comparisons; they are not independent evidence.

Equivalent symbolic quotients are canonicalized before distinct values are counted. For the representation-information tariff the modal value is 1/(8-3*log2(3))≈0.308155726702, occurring 17,100 times. The raw symbolic export contains 37 textual quotient forms; exact algebraic normalization collapses these to 17 distinct numerical threshold values. Thus “17” counts canonical values, whereas “37” counts pre-canonical symbolic spellings.

General boundary propositions

  1. Refinement monotonicity. If a deterministic representation r_f strictly refines r_c and both optimize over all deterministic latent-conditioned policies, then V∗(rf)≥V∗(rc)V^*(r_f) \ge V^*(r_c). Proof: compose any policy on r_c with the coarsening map from r_f; the finer policy class contains an imitation of every coarser policy. For the broader comparison-of-experiments framework, see Blackwell [3], listed in the manuscript references. The argument here is the explicit deterministic policy-composition proof above, with no additional assumptions imported.

  1. Binary complement symmetry. Suppose actions are binary, the decision-code library is closed under bitwise complement, both identity and bit-flip action codes are present, and payoff is symmetric zero-one correctness. Then each realized action map has at least two encodings. Paired fixed-background payoff contrasts have opposite signs while representation information is unchanged. The sign-pairing statement is not asserted outside this domain.

  1. Census non-invariance. Let phenotype ii have encoding multiplicity mim_i. A census that includes all unordered encoded pairs contains ∑i(mi2)\sum_i \binom{m_i}{2} within-phenotype pairs and ∑i<jmimj\sum_{i<j}m_i m_j between-phenotype pairs. Uniformly multiplying every mim_i by the same factor does not generally preserve counts or class shares because the two terms scale differently. If within-phenotype pairs are excluded and a class depends only on the phenotype pair, uniform duplication multiplies every class by the same squared factor and shares are invariant. Therefore every census must state both its pair universe and its counting measure.

  1. Information-tariff identity. For C=I(W;R)C=I(W;R), a pair has a positive threshold exactly when its zero-cost payoff leader has higher retained world information. This follows directly from λ∗=ΔV/ΔC>0\lambda^*=\Delta_V/\Delta_C>0; it is an algebraic identity, not independent empirical support.

These propositions are conditional on the policy and coding assumptions stated above. The first, the explicitly conditional third, and the fourth have forms that apply beyond the particular four-state counts. The second is exact only in the binary, complement-closed, symmetric-payoff domain just stated. None is a biological theorem.

Tie and null structure

All 1,267 module/environment optima are set-valued. Stable tasks have optimum-set sizes 2, 4, or 8; two-task mixtures additionally have size 16; the all-task mixture has sizes 2 or 4. No singleton optimum can exist because every realized action map has a decision-complement/action-flip encoding partner. Neutral encodings are retained in raw and aggregate artifacts; their existence is not interpreted as an evolutionary result.

Likewise, the 792:792 strict representation-replacement count is forced by the action-complement pairing. Each paired background preserves information order and negates payoff difference. The existence of both strict directions is empirical to the finite atlas; their exact equality is a model invariant.

Null and opposite-direction results are part of the main evidence: equal-information comparisons, unequal-information payoff ties, strict higher-information/higher-payoff comparisons, and tariff classes without positive crossings are not filtered.

Independent validation

Fresh-context critical reviews accepted all three executions with zero findings. Their reports record blind first-principles reconstruction, and repository checker evidence is bound where it exists. The validated scope was:

  • the complete sets of Stage 7.2A raw and aggregate results;

  • all 44,775 Stage 7.2B scientific rows;

  • all 916,265 Stage 7.2C scientific rows, symbolic threshold signs, tie sets, and aggregates.

The Stage 7.2B and 7.2C repository checker implementations and oracle evidence declare no import of their production scientific cores. The Stage 7.2A review report records a standalone blind reconstruction; its compact oracle evidence is separately bound. Publication figures and robustness tables are generated from accepted saved raw/aggregates by python-mfts_lab.publication.paper_001.

Claim boundary

Allowed:

Within the complete locked uniform four-state model, lower retained world information can accompany higher expected task payoff, but the reverse relation and ties also occur. Their zero-cost census changes with the comparison universe and registered task mixture, while separately declared mechanism tariffs change pairwise J_lambda net-performance ordering.

This is the content-addressed claim-ledger wording. In ordinary journal language, “complete locked” means completely enumerated within the finite S2 domain, and “registered task mixture” means one of the seven declared stationary objectives. It does not imply public preregistration.

Not supported:

  • I(W;R)I(W;R) is truth simpliciter;

  • task payoff or JλJ_\lambda is reproductive fitness;

  • pairwise order is invasion, fixation, extinction, or prevalence;

  • contextual locus contrasts are context-free causal allele effects;

  • project tariffs are empirical costs;

  • finite counts are probabilities over agents;

  • the model represents a named organism, human perception, or arbitrary ML systems;

  • the results prove FBT, ITP, consciousness, spacetime, ontology, or metaphysics.

AI-assisted workflow disclosure

OpenAI Codex assisted with source organization, implementation, drafting, and fresh-context review. The Human Principal Investigator approved the scientific assumptions, byte-exact configurations, publication claim, and internal lifecycle. Deterministic oracles and isolated verifier runs checked artifacts. The first author reviewed scientific decisions and interpretations and retains responsibility for the manuscript. Internal AI-assisted checks are not external human peer review. AI systems are not listed as authors.

Artifact map

The accompanying PAPER-001-code-and-data.zip is an article-scoped public package. After extraction, its README.txt and package_manifest.json identify all files:

  • current editable manuscripts and supplement: publications/paper-001/;

  • formal definitions and exact fixtures: specs/integrated/;

  • unchanged exact configurations: experiments/configs/;

  • unchanged raw and derived results: results/raw/ and results/derived/;

  • original computation manifests: experiments/manifests/;

  • figures: publications/paper-001/figures/ and figures_ru/;

  • figure and supplementary-table inputs: publications/paper-001/source_tables/;

  • figure checksums and input bindings: publications/paper-001/figure_manifest.json;

  • claim-to-evidence trace: publications/paper-001/claim_evidence_ledger.csv;

  • scientific core, independent checkers, and saved-data figure builder: src/fts_lab/;

  • portable verification and reproduction commands: README.txt, verify_package.py;

  • pinned verification dependencies: requirements-verification.txt;

  • original project dependency files: environment/ (historical provenance).

Machine-readable table filenames retain their original identifiers. Current supplementary table numbers map to them as follows:

  • Table S3: table_s5_information_decomposition_witness.csv;

  • Table S11: table_s1_representation_once_census.csv;

  • Table S12: table_s2_nested_refinement_control.csv;

  • Table S14: table_s3_one_coordinate_diagnostics.csv;

  • Table S15: table_s7_worked_order_examples.csv;

  • Table S17: table_s4_threshold_distributions.csv.

Historical review documents named by original manifests or the claim ledger are not distributed here. They record internal AI-assisted checking, not external human peer review. The public verification commands do not require those files. Original computational provenance, including generic workspace paths and Git identifiers, is retained to preserve raw-file identity. The original non-executable configuration flags remain unchanged; see README.txt for the public verification workflow. Primary-source papers and private correspondence are not included.

References

  1. [1]Angelucci, A., Fano, V., Ferretti, G., Macrelli, R., & Tarozzi, G. (2021). Evolutionary dynamics and accurate perception: Critical realism as an empirically testable hypothesis. Philosophia Scientiæ, 25-2, 157–178. https://doi.org/10.4000/philosophiascientiae.2960
  2. [2]Berke, M. D., Walter-Terrill, R., Jara-Ettinger, J., & Scholl, B. J. (2022). Flexible goals require that inflexible perceptual systems produce veridical representations: Implications for realism as revealed by evolutionary simulations. Cognitive Science, 46(10), e13195. https://doi.org/10.1111/cogs.13195
  3. [3]Blackwell, D. (1953). Equivalent comparisons of experiments. The Annals of Mathematical Statistics, 24(2), 265–272. https://doi.org/10.1214/aoms/1177729032
  4. [4]Hoffman, D. D. (2019). The case against reality: Why evolution hid the truth from our eyes. W. W. Norton & Company.
  5. [5]Mark, J. T., Marion, B. B., & Hoffman, D. D. (2010). Natural selection and veridical perceptions. Journal of Theoretical Biology, 266(4), 504–515. https://doi.org/10.1016/j.jtbi.2010.07.020
  6. [6]Martínez, M. (2019). Usefulness drives representations to truth: A family of counterexamples to Hoffman’s interface theory of perception. Grazer Philosophische Studien, 96(3), 319–341. https://doi.org/10.1163/18756735-09603004
  7. [7]Marzen, S. E., & DeDeo, S. (2017). The evolution of lossy compression. Journal of the Royal Society Interface, 14(130), 20170166. https://doi.org/10.1098/rsif.2017.0166
  8. [8]Mattes, J. (2026). The case against The case against reality. Biological Theory. https://doi.org/10.1007/s13752-026-00531-0
  9. [9]Prakash, C., Stephens, K. D., Hoffman, D. D., Singh, M., & Fields, C. (2021). Fitness beats truth in the evolution of perception. Acta Biotheoretica, 69, 319–341. https://doi.org/10.1007/s10441-020-09400-0
  10. [10]Prakash, C., Fields, C., Hoffman, D. D., Prentner, R., & Singh, M. (2020). Fact, fiction, and fitness. Entropy, 22(5), 514. https://doi.org/10.3390/e22050514
  11. [11]Rauh, J., Banerjee, P. K., Olbrich, E., Jost, J., Bertschinger, N., & Wolpert, D. (2017). Coarse-graining and the Blackwell order. Entropy, 19(10), 527. https://doi.org/10.3390/e19100527
  12. [12]Shannon, C. E. (1948). A mathematical theory of communication. Bell System Technical Journal, 27(3–4), 379–423, 623–656. July installment: https://doi.org/10.1002/j.1538-7305.1948.tb01338.x October installment: https://doi.org/10.1002/j.1538-7305.1948.tb00917.x
  13. [13]Strouse, D. J., & Schwab, D. J. (2017). The deterministic information bottleneck. Neural Computation, 29(6), 1611–1630. https://doi.org/10.1162/NECO_a_00961
  14. [14]Tan, A. K., Tegmark, M., & Chuang, I. L. (2022). Pareto-optimal clustering with the primal deterministic information bottleneck. Entropy, 24(6), 771. https://doi.org/10.3390/e24060771
  15. [15]Tegmark, M., & Wu, T. (2020). Pareto-optimal data compression for binary classification tasks. Entropy, 22(1), 7. https://doi.org/10.3390/e22010007
  16. [16]Venkataramanan, S., Pariza, V., Salehi, M., Knobel, L., Ramzi, E., Gidaris, S., Bursuc, A., & Asano, Y. M. (2026). Franca: Nested Matryoshka clustering for scalable visual representation learning. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10533–10544. https://openaccess.thecvf.com/content/CVPR2026/papers/Venkataramanan_Franca_Nested_Matryoshka_Clustering_for_Scalable_Visual_Representation_Learning_CVPR_2026_paper.pdfarxiv.org/abs/2507.14137

Paper details

vlnc:W_9f9cqx9x
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Codex

View statement
Looking for related papers…
Contents