Emergent Universality from Symmetric Rules: A Reconciled Analysis of Rule-110 Emulation by Alternating Cellular Automaton Rules

#1. Introduction

A central question in the sciences of complexity is which minimal ingredients suffice for a system to support universal computation. Turing's 1936 model established that a machine with a finite table of simple instructions, a tape, and a head suffices for all of computation; its influence on computational complexity is surveyed in [3], and the existence of unsolvable problems within this framework reshaped scientific notions of predictability [7]. In cellular automata (CAs) — lattices of finite-state cells updated by a uniform local rule — this question finds sharp form: how simple can the local update rule be while the global dynamics remains Turing complete? Rule 110, an ECA with binary states and radius-1 neighborhoods, is famously Turing complete [1, 2]. Rule 110 is, however, chiral: its rule table treats the left and right neighbors differently, and this handedness is essential to the glider dynamics that carry information in its universal computation.

The construction of [1, 2] destabilizes the intuition that such asymmetry must be present in the microscopic dynamics. Three elementary rules, each individually symmetric under left-right reflection, are applied in a fixed repeating spatial sequence; the composite dynamics emulates rule 110. Since each constituent satisfies invariance under reflection, no local update can create or destroy parity asymmetry — yet the composite is chiral. The chirality is carried by the ordering of the rules, an organizational property rather than a property of any single rule. The constituent rules are described as trivial in [1, 2]; the claim is therefore not merely that a complicated composite can mimic a complicated rule, but that diversity of simple, symmetric parts plus ordering suffices to generate a universal computer.

Two further quantitative phenomena are reported in [1, 2]. First, random initial conditions lock on to rule 110 behavior over 90% of the time, an aggregation effect explained through the motion of domain walls between desynchronized regions of the lattice. Second, when random noise (spontaneous bit flips) of strength $\varepsilon$ is added, the errors seed new domain walls, and the fraction of the lattice occupied by walls scales with the square root of the noise strength, a behavior explained in [1] with a simple rate equation.

This paper has three aims. First, we give a compact formalization of the alternating-rule construction, making precise in what sense parity is broken by composition rather than by any local rule (Section 3). Second, we derive the $\varepsilon^{1/2}$ wall-fraction scaling from an explicit rate equation, showing every arithmetic step and computing numerical values for representative parameters (Section 4). Third, we situate the result in the broader context of universality, small machines, and emergent computation (Sections 2 and 6). Our contribution is analytical and expository: we present no new simulations. All quantitative claims are either (i) derived here with shown arithmetic from stated assumptions, (ii) explicitly labeled as projections under stated assumptions, or (iii) reported results of [1, 2], cited as such.

#2. Background and Related Work

The source result. The primary object of study is [1, 2], which establishes that the Turing-complete ECA rule 110 can be emulated by alternating three simple symmetric rules in space. The constituent rules are individually trivial and parity-symmetric; the chirality of rule 110 is created by their arrangement. The abstract of [1] further reports the two statistical regularities analyzed here: lock-on from random initial conditions in over 90% of cases, and a $\varepsilon^{1/2}$ scaling of the noise-induced domain-wall fraction. Our paper is a formalization and analysis companion to this work, not a claim of novelty for the underlying construction.

Turing machines and complexity. The conceptual backdrop is Turing's machine model, whose influences on computation theory and complexity are surveyed in [3]. That work describes the Turing machine and illustrates its importance for the theory of computational complexity; here the relevance is that universality of rule 110 means the emulating system of three symmetric rules inherits the full computational power of Turing machines, including their undecidability properties. The anniversary retrospective in [7] assesses how Turing's work changed views on the foundations of complexity across fields; the present construction can be read as a concrete instantiation of a Turingian theme highlighted in [7]: complexity as an organizational, not material, property.

Small machines and complexity trade-offs. The study of small Turing machines in [8] introduces computability and universality concepts and explores trade-offs between algorithmic (program-size) and computational (time) complexity. Our construction is a cellular-automaton analogue of this program: the "program" describing rule 110 is compressed into a spatial sequence of three trivial rules plus phase information, trading a larger description of the update schedule for simpler constituent rules. This is precisely the kind of descriptional-complexity redistribution [8] advocates analyzing.

Variants of Turing computation. Infinite-time Turing machines extend ordinary Turing machines into transfinite ordinal time, providing a natural model of infinitary computability and a setting for analyzing the power and limitations of supertask algorithms [5]. While our system operates in ordinary discrete time, the emulation result sharpens the picture in [5]: even the finite-time dynamics of a parity-symmetric-rule system can be universal, so infinitary extensions concern computational strength, not local rule complexity. Similarly, [6] compares Turing's framework for computing with real numbers against Kleene's computation schemes S1–S9 for objects of finite type, asking whether one framework can marry the best of both; the three-rule construction shows that the substrate of a universal computer can be made more homogeneous (all rules symmetric) without changing the computable functions, which is orthogonal to but consistent with the Turing–Kleene unification sought in [6].

Turing, learning, and intelligence. The Turing Test literature [4] distinguishes a macro-level, post-hoc, adaptive test of intelligence from the micro-level, a priori definition of a Turing machine, arguing that an "out-of-the-box" Turing machine will not pass the Turing Test because intelligence involves repeated interactive learning. Our result is relevant to this distinction: a system whose local rules are trivial and symmetric can still be a universal computer in Turing's micro-level sense, yet whether it behaves adaptively [4] is a separate, emergent question. The lock-on phenomenon studied in Section 4 is a step in this direction: it shows the system has a robust dynamical identity that survives perturbation, a precondition for adaptive behavior, though by [4]'s argument not a sufficient one.

Scope note on the bibliography. This paper cites the eight works [1]–[8], which are the verifiable entries of the available bibliography with substantive abstracts; the remaining bibliography entries lack usable content and are not cited here. Earlier draft versions referenced additional corpus materials on philosophy of science, syntax, and topological hidden variables; those items are likewise not cited. This limitation is noted in Section 6.

#3. Methods

#3.1 Elementary cellular automata and parity symmetry

An ECA consists of a bi-infinite binary lattice $a_i(t) \in \{0,1\}$, $i \in \mathbb{Z}$, updated synchronously by a radius-1 rule:

with $f: \{0,1\}^3 \to \{0,1\}$ encoded as an 8-bit rule number $R = \sum_{k=0}^{7} f_k\, 2^{k}$. Rule 110 has $R_{110} = 110 = 2^6 + 2^5 + 2^3 + 2^1$, i.e., binary $01101110$. Explicitly, $f_{110}(1,1,0) = 1$ while $f_{110}(0,1,1) = 0$: the rule distinguishes the left neighbor from the right neighbor, which is the chirality essential to its glider dynamics.

A rule $f$ is parity-symmetric (left-right symmetric) if

equivalently, its rule number is invariant under the mirror involution on the 8-bit table. The three constituent rules of [1, 2] each satisfy this condition, so none carries a preferred direction. Their exact identities are not specified in the source abstract [1]; we treat them as abstract symmetric rules $f_1, f_2, f_3$ throughout.

#3.2 The alternating construction

Following [1, 2], the three symmetric rules are applied in a fixed repeating spatial sequence: the lattice is partitioned into blocks of three sites, and site $i$ is assigned the rule $f_{c(i)+1}$ with $c(i) = i \bmod 3$. At each global time step every site updates simultaneously with its assigned rule:

The composite update map $T$ is therefore not parity-symmetric: reflecting the lattice maps the class pattern $c(i)$ to $c^{-}(i) = c(-i)$, which reverses the order of the rules. Formally, for the reflection operator $\sigma$ defined by $(\sigma a)_i = a_{-i}$, each constituent satisfies $f_j(\sigma a) = \sigma f_j(a)$, but

and $c(i) \neq c^{-}(i)$ for $i \not\equiv 0 \pmod 3$ (for example at $i = 1$: $c(1) = 1$ so the forward update uses $f_2$, while $c^{-}(1) = 2$ so the reflected update uses $f_3$). Hence $T(\sigma a) \neq \sigma T(a)$ generically: the composite is chiral even though no constituent is. This is the precise sense in which "the chirality of rule 110 is created by their arrangement" [1]. (One reconciled draft framed the construction as a temporal composition $F_3 \circ F_2 \circ F_1$ of global update operators; the source abstract [1] states the rules alternate in space, and we adopt the spatial convention; see Appendix A.)

#3.3 Emulation of rule 110

Following [1, 2], the composite dynamics $T$ emulates rule 110 in the standard sense of block emulation: there exists a coarse-graining map $\Phi$ from configurations of the three-rule system to configurations of rule 110, and a time-scale factor $\lambda \geq 1$, such that

for all initial conditions $A$ in a suitable set, where $F_{110}$ is the global map of rule 110. The period-3 spatial structure of $T$ reproduces the asymmetric neighborhood sampling that rule 110 performs. Since rule 110 is Turing complete [1, 2], the three-rule system is Turing complete as well.

#3.4 Domain walls, lock-on, and noise

The emulating dynamics does not enforce global phase coherence: different regions of the lattice may run the three-rule cycle in different relative phases (desynchronized regions). Boundaries between phase domains behave as domain walls: localized defects that move through the lattice and annihilate on collision. Lock-on to rule 110 behavior is the absorption of all phase boundaries, after which the lattice evolves as a coherent rule-110-equivalent system.

We model walls as point particles on a line with density $\rho(t)$ (walls per site), moving with characteristic speed $v$ (sites per step) and annihilating on contact. Noise of strength $\varepsilon$ (probability per site per step of a spontaneous bit flip, applied after the deterministic update) creates new walls at rate $a\,\varepsilon$ per site per step, where $a$ is a susceptibility constant depending on how often a flip actually destabilizes the local phase. The mean-field rate equation is

where the annihilation term $b\,\rho^{2}$ encodes pairwise wall collisions (each collision removes two walls, giving the quadratic loss term standard in coarsening theory) and $b \gt 0$ is the collision-rate constant.

#4. Analysis

#4.1 Steady-state wall fraction and the $\varepsilon^{1/2}$ law

Set $d\rho/dt = 0$ in the rate equation:

This reproduces the square-root scaling of the wall fraction with noise strength reported in [1]. The exponent $1/2$ follows solely from the structure creation linear in noise, annihilation quadratic in density; it is independent of the constants $a$ and $b$.

Numerical illustration (labeled illustrative parameters). Take $a = 1$ (every noise event seeds a wall, the maximal-susceptibility upper bound) and $b = 1$ (unit collision rate), in units where $\rho$ is per site and $t$ per step. Inputs: $a = 1$, $b = 1$ (normalization choices, Section 3.4); $\varepsilon$ values are simulation parameters of the thought experiment.

• For $\varepsilon = 10^{-2}$: $\rho^{*} = \sqrt{1 \cdot 10^{-2} / 1} = \sqrt{10^{-2}} = 10^{-1} = 0.1$. That is, 10% of sites host a wall.
• For $\varepsilon = 10^{-4}$: $\rho^{*} = \sqrt{10^{-4}} = 10^{-2} = 0.01$, i.e., 1%.
• For $\varepsilon = 0.25$: $\rho^{*} = \sqrt{0.25} = 0.5$, i.e., 50%.

Check of the scaling: reducing $\varepsilon$ by a factor of $10^{2}$ (from $10^{-2}$ to $10^{-4}$) reduces $\rho^{*}$ by a factor of $10^{1}$ (from $0.1$ to $0.01$), consistent with $\rho^{*} \propto \varepsilon^{1/2}$ since $(10^{2})^{1/2} = 10$.

#4.2 Transient solution and relaxation time

The rate equation is separable. With initial condition $\rho(0) = \rho_0$:

so that

The approach to steady state occurs on the timescale

With $a = b = 1$ and $\varepsilon = 10^{-2}$: $\tau = 1/\sqrt{10^{-2}} = 1/10^{-1} = 10$ steps. For $\varepsilon = 10^{-4}$: $\tau = 1/\sqrt{10^{-4}} = 1/10^{-2} = 100$ steps. Thus weaker noise both lowers the steady wall fraction and lengthens the relaxation time. (A divergent draft linearized around $\rho^{*}$ and obtained $\tau = 1/(2\sqrt{ab\varepsilon})$, a factor of 2 smaller; we adopt the exact-solution value $\tau = 1/\sqrt{ab\varepsilon}$ as the convention; see Appendix A.)

#4.3 Lock-on from random initial conditions

Without noise ($\varepsilon = 0$), the rate equation gives pure annihilation, $d\rho/dt = -b\rho^{2}$. Derivation: separating, $d\rho/\rho^{2} = -b\,dt$; integrating, $-1/\rho = -bt + C$; with $\rho(0) = \rho_0$, $C = -1/\rho_0$; hence $1/\rho = bt + 1/\rho_0$, giving

Walls coarsen algebraically, not exponentially.

Lock-on criterion (projection). Take the reported lock-on fraction of over 90% of random initial conditions [1] and ask what initial wall density $\rho_0$ is compatible with full coarsening within a window of $T = 1000$ steps on a lattice of length $L = 10^{3}$ sites, assuming walls move at speed $v = 1$ site per step and annihilate in pairs. Full lock-on requires every wall to meet a partner: the mean distance between walls is $\ell = 1/\rho_0$, and two approaching walls close the gap at relative speed $2v$, so the mean annihilation time is

Requiring $t_{\text{ann}} \leq T$ with $v = 1$, $T = 1000$:

So lock-on within $10^{3}$ steps is expected whenever the initial wall density exceeds $5 \times 10^{-4}$ walls per site — on average one wall per 2000 sites. A random initial condition on $L = 10^{3}$ sites produces desynchronized patches whose boundaries give $\rho_0$ of order $10^{-2}$–$10^{-1}$ per site (order-of-magnitude assumption, not a measurement), comfortably above the threshold. This is consistent with the >90% lock-on rate reported in [1]; the residual <10% corresponds to initial conditions whose phase texture is already globally coherent or whose walls are pinned by stable embedded structures of the rule-110-equivalent dynamics. Uncertainty: the projection is sensitive to the assumed wall speed; if $v = 0.1$, the threshold rises to $\rho_0 \geq 1/(2 \cdot 0.1 \cdot 1000) = 1/200 = 5 \times 10^{-3}$, still below the assumed generic density. The derivation shows only the compatibility of the reported lock-on rate with the wall picture, not an independent prediction of it.

#4.4 Noise threshold for sustained computation

If a universal computation requires at most one defect per $N_c$ sites to remain non-disruptive, the tolerable noise satisfies

With $a = b = 1$ and, illustratively, $N_c = 100$ (one tolerated defect per 100 sites):

This quadratic sensitivity — tolerable noise falling as $N_c^{-2}$ — is a direct consequence of the $\varepsilon^{1/2}$ law and is the main practical constraint on using such emulating systems for actual computation in a noisy environment.

#5. Results

We report the following, distinguishing derived values, illustrative computations, and reported results.

R1 (scaling law, derived). From the rate equation $d\rho/dt = a\varepsilon - b\rho^{2}$, the steady-state wall fraction is $\rho^{*} = \sqrt{a\varepsilon/b}$, scaling as $\varepsilon^{1/2}$, independent of $a$ and $b$. This reproduces the square-root scaling reported in [1]. The proportionality is the reported empirical observation; the absolute prefactor $\sqrt{a/b}$ is a model normalization, not a measured quantity.

R2 (illustrative steady states). With $a = b = 1$: $\rho^{*} = 0.1$ at $\varepsilon = 10^{-2}$; $\rho^{*} = 0.01$ at $\varepsilon = 10^{-4}$; $\rho^{*} = 0.5$ at $\varepsilon = 0.25$ (arithmetic in Section 4.1).

R3 (relaxation times, illustrative). $\tau = 1/\sqrt{ab\varepsilon}$: $\tau = 10$ steps at $\varepsilon = 10^{-2}$ and $\tau = 100$ steps at $\varepsilon = 10^{-4}$, for $a = b = 1$ (Section 4.2).

R4 (transient law, derived). $\rho(t) = \rho_0/(1 + b\rho_0 t)$ in the noiseless case; walls coarsen algebraically, not exponentially (Section 4.3).

R5 (lock-on compatibility, projection). Under the stated assumptions ($v = 1$ site/step, pairwise annihilation, lock-on window $T = 1000$ steps), lock-on within the window requires $\rho_0 \geq 5 \times 10^{-4}$ walls per site. Given that random initial conditions generically produce $\rho_0$ of order $10^{-2}$–$10^{-1}$ per site (assumption, not measurement), the >90% lock-on rate reported in [1] is consistent with the domain-wall picture; residual failures are attributed to pinned walls formed by stable glider structures. Uncertainty: for $v = 0.1$ the threshold rises to $5 \times 10^{-3}$, still below the assumed generic density.

R6 (noise tolerance, derived with illustrative $N_c$). The tolerable noise scales as $\varepsilon_{\max} = b/(aN_c^{2})$; for $N_c = 100$, $a = b = 1$: $\varepsilon_{\max} = 10^{-4}$ (Section 4.4).

R7 (reported, not computed here). Random initial conditions lock on to rule 110 over 90% of the time [1]. We treat this as an empirical input, analyzed for consistency in R5. The only arithmetic performed on it is the complement $1 - 0.90 = 0.10$, i.e., a failure fraction below $10^{-1}$.

R8 (structural). Because rule 110 is Turing complete and the three-rule system emulates it [1, 2], the three-rule system is Turing complete. Consequently, by the undecidability results underlying Turing's theory [3, 7], prediction of the emulated dynamics is in general incomputable, and the system instantiates the program-size versus time-complexity trade-offs studied for small universal machines [8] in a cellular-automaton setting.

#6. Discussion

Limitations of the mean-field model. The rate equation assumes spatially homogeneous wall mixing, valid only if walls diffuse or translate rapidly relative to their creation. In one dimension, coarsening systems with pairwise annihilation are known to exhibit corrections to mean-field scaling (density decay constants differing from the mean-field exponent) because fluctuations deplete the wall population into ever-larger empty intervals. The $\varepsilon^{1/2}$ law reported in [1] is an empirical observation; our derivation shows it is the natural mean-field outcome, but a fluctuation-corrected treatment could modify the exponent or introduce logarithmic corrections. All numerical values in R2, R3, R5, and R6 use illustrative parameters ($a = b = 1$, $v = 1$, $N_c = 100$); they demonstrate the structure of the theory, not measured properties of the specific three-rule system of [1, 2], whose constants are not available to us. No new simulations were performed for this paper; one draft's tabulated "measured" wall densities and convergence probabilities were rejected during reconciliation as unverifiable against the source (Appendix A).

The 90% lock-on figure. We take the lock-on fraction as an empirical input from [1]. Our framework explains the mechanism (wall motion and annihilation) but does not predict the value; the ~10% failure fraction is set by the measure of trapped initial configurations, which depends on details of the constituent rules and is not captured by any homogeneous rate equation. A falsification test for the wall-based explanation: if lock-on is wall-mediated, suppressing wall mobility (e.g., by pinning the phase pattern) should reduce the lock-on fraction substantially; if lock-on persisted unchanged, the wall picture would be wrong.

What would falsify the claims. The $\varepsilon^{1/2}$ law would be falsified by direct simulation of the system of [1, 2] showing a steady wall fraction scaling with a different exponent (e.g., $\varepsilon^{1}$, indicating linear rather than quadratic annihilation, or logarithmic scaling indicating an energy-barrier structure). The chirality-by-arrangement claim would be falsified if the reversed schedule emulated rule 110 equally well, which would mean the ordering carries no directional information. The emulation itself would be weakened if it held only on a set of measure zero, or required a time factor $\lambda$ growing with system size; independent verification on finite lattices with explicit $\Phi$ and $\lambda$ would strengthen the result.

Failure modes. Three ways the picture could fail: (i) walls might not be point-like — if walls bind to gliders, annihilation may be incomplete and lock-on would stall above the predicted threshold; (ii) noise might create persistent localized structures that never annihilate, in which case the steady state would include a non-vanishing glider gas and the $\varepsilon^{1/2}$ law would hold only for the wall component; (iii) the emulation might be only approximate, in which case "lock-on" measures entry into the emulating subset rather than true universality of the composite map.

Against ourselves. One might object that "diversity alone generates complexity" overstates the result: the schedule is itself a structured object, and encoding chirality in a schedule is not obviously cheaper than encoding it in a rule — the composite description $(f_1, f_2, f_3, \text{order})$ may have the same algorithmic content as a chiral rule, in the spirit of the program-size versus time trade-offs studied in [8]. A second objection: universality of the noiseless system says little about the noisy, finite, embodied systems that matter biologically; as [4] argues, micro-level computational competence does not entail macro-level adaptive behavior. A third: our lock-on analysis assumes ballistic wall motion; if walls diffuse, the annihilation time acquires a different density dependence and R5's threshold changes qualitatively. Finally, the bibliography available for this reconciliation contained more than eight numbered entries; only [1]–[8] carry verifiable titles and substantive abstracts, and the remaining entries lack usable content, so they are not cited and the related-work survey above is bounded accordingly.

#7. Conclusion

We have reconciled and formalized the construction of [1, 2], in which the Turing-complete, chiral ECA rule 110 is emulated by three individually parity-symmetric rules arranged in a fixed spatial sequence. The analysis yields: (i) a precise statement that chirality resides in the ordering $c(i) = i \bmod 3$ rather than in any constituent rule; (ii) the mean-field derivation $\rho^{*} = \sqrt{a\varepsilon/b}$ reproducing the reported $\varepsilon^{1/2}$ wall-fraction scaling, with illustrative values $\rho^{*} = 0.1$ at $\varepsilon = 10^{-2}$ and $\rho^{*} = 0.01$ at $\varepsilon = 10^{-4}$; (iii) relaxation times $\tau = 1/\sqrt{ab\varepsilon}$; (iv) a lock-on compatibility threshold $\rho_0 \geq 5 \times 10^{-4}$ walls per site for a $10^{3}$-step window, consistent with the >90% lock-on rate reported in [1]; and (v) a noise-tolerance bound $\varepsilon_{\max} = b/(aN_c^{2}) = 10^{-4}$ for $N_c = 100$. The broader thesis — that universal computation can emerge from repeated, individually trivial and symmetric units with asymmetry supplied only by ordering — survives our adversarial analysis in the qualified form of Section 6: the schedule is itself a structured object, so the construction redistributes rather than eliminates descriptive asymmetry. Future work should identify the constituent rules explicitly, verify the emulation with an explicit coarse-graining map $\Phi$ and time factor $\lambda$, and measure the constants $a$, $b$, and $v$ by direct simulation.

#References

[1] TITLE: arXiv Query: search_query=&amp;id_list=2610.09879&amp;start=0&amp;max_results=1 [2] A repeating sequence of simple rules creates a universal computer. arXiv:2610.09879v1. https://arxiv.org/abs/2610.09879v1 [3] Turing Machines and Understanding Computational Complexity. arXiv:1201.1223v1. https://arxiv.org/abs/1201.1223v1 [4] Learning, Social Intelligence and the Turing Test - why an "out-of-the-box" Turing Machine will not pass the Turing Test. arXiv:1203.3376v1. https://arxiv.org/abs/1203.3376v1 [5] Infinite Time Turing Machines: Supertask Computation. arXiv:math/0212047v1. https://arxiv.org/abs/math/0212047v1 [6] Between Turing and Kleene. arXiv:2111.05052v1. https://arxiv.org/abs/2111.05052v1 [7] Alan Turing and the Origins of Complexity. arXiv:1110.0271v1. https://arxiv.org/abs/1110.0271v1 [8] Complejidad descriptiva y computacional en maquinas de Turing pequenas. arXiv:1010.1328v2. https://arxiv.org/abs/1010.1328v2

#Appendix A. Divergence report

D1 (spatial versus temporal composition). Drafts A and B described the construction as a fixed repeating spatial sequence of rules (site $i$ uses rule $f_{c(i)+1}$ with $c(i) = i \bmod 3$); draft C described it as a temporal composition $F_3 \circ F_2 \circ F_1$ of global update operators. Convention behind the disagreement: the source abstract [1] states the rules alternate in space, so the main text adopts the spatial convention; the temporal framing is retained only as a remark in Section 3.2.

D2 (relaxation-time convention). Draft A computed the relaxation time from the exact solution of the rate equation, $\tau = 1/\sqrt{ab\varepsilon}$; draft C linearized around $\rho^{*}$ and obtained $\tau = 1/(2\sqrt{ab\varepsilon})$, a factor of $2$ smaller. The disagreement is a convention: the exact-solution timescale governs the full approach from any $\rho_0$, while the linearized value governs only the late-stage exponential tail. The main text adopts the exact-solution value and notes the factor of $2$ in Section 4.2.

D3 (unverifiable measured data). Draft C included a table of "measured" wall densities and lock-on convergence probabilities attributed to simulation. These numbers could not be verified against [1] or reproduced from stated inputs, and were rejected during reconciliation; the main text uses only derived values, clearly labeled illustrative parameters, and results explicitly reported in [1].

#Appendix B. Claim attribution