NOTE-1493: Baumgartner & Gebharter — fat-handedness

Contribution

It shows that Craver’s mutual manipulability criterion (MM), read with the interventionist definitions it was built on, cannot do the job it was built for. Under Woodward 2003, MM plus supervenience plus non-causal constitution entails that no part is ever constitutively relevant to a whole (pp. 10–13). Under Woodward’s 2011 relaxation it entails that whole and part are causally related (pp. 15–17). A temporal clause on causation removes both absurdities (pp. 17–18). Then a new problem appears: every mutual manipulation is a common cause of whole and part, so the manipulations give no evidence of constitution (pp. 18–20). The paper closes with a second-order, abductive criterion, fat-handedness (FH).

Key insight

Since a whole cannot change without some part changing, nothing can act on the whole except by also acting on a part, along a different path. Every “top-down” manipulation is a common cause of whole and part. A common cause of two variables tells you nothing about how they are related to each other. So mutual manipulability, which looks like the experimental mark of constitution, is equally well explained by “the experimenter wiggles both” (Fig. 5b). What a constitutive model explains and the common-cause model does not is a pattern across experiments: no cause of the whole is ever found that leaves every part alone (pp. 20–21).

Assumptions

• (Sup) “Every change in the supervening behavior is necessarily accompanied by a change in the supervenience base” (p. 5). The base is taken to be the constituents’ behaviours, not their causal organization, which under a difference-making theory adds nothing (pp. 5–6).
• (SC) constituents are spatiotemporal parts of the macro level; (NR) a part that never makes a difference is not a constituent (p. 6).
• (NC) constitution is non-causal (p. 7). The majority view, assumed standardly, with the consequences of dropping it examined (pp. 13–15).
• Stability. The mechanism is “completely stable”, with compensatory responses cancelled by further interventions or the setup (p. 10).
• Variables as modelling devices. The ontology of macro and micro levels is left open; Ψ and the Φᵢ are “different random variables such that changes in the values of Ψ supervene on changes in V1” (p. 6).
• Parthood is not analysed. (SC) is used as given; its problems are referred to Leuridan’s §6 (LIT-1929) and Franklin-Hall (p. 6).

Key results

• MM as stated (p. 8). Φᵢ(Xᵢ) is constitutively relevant to Ψ(S) iff Xᵢ is a spatiotemporal part of S, some possible ideal intervention on Φᵢ changes Ψ, and some possible ideal intervention on Ψ changes Φᵢ. The authors read it as necessary and sufficient, against Couch (footnote 4).
• (INEX) under Woodward 2003 (pp. 10–13). Suppose I_Ψ is an intervention variable for Ψ w.r.t. Φᵢ that changes both. By (IV.i) and (M), I_Ψ causes Ψ; by (Sup), every change it induces in Ψ comes with a change in some Φⱼ, so I_Ψ causes Φⱼ too. Of the three structures in which one variable causes two (Fig. 3), (NC) leaves only the common cause (C). But (IV.iii) requires every path from I_Ψ to Φᵢ to pass through Ψ. So either I_Ψ is not an intervention, or it does not change Φᵢ. Condition (†), that each Φᵢ be changeable by a top-down intervention, is unsatisfiable “for any mechanism” (p. 13). Hence: “There do not exist any relationships of constitutive relevance between macro phenomena and their spatiotemporal parts.” Checked: the argument uses only (Sup), (NC), (M) and (IV), and fails if (NC) is dropped, as the authors say.
• Dropping (NC) (pp. 13–15). Ψ and Φᵢ can then sit on one path, (†) becomes satisfiable, and MM yields “constitutive causation”: bidirectional, instantaneous, whole-to-part causal loops (Fig. 4). This is Leuridan’s conclusion (his p. 424), which the authors grant follows but reject as unlike any causal structure.
• Under Woodward 2011, interventionism (pp. 15–17).* (IV*) lets a path from I to Y bypass X if it passes through X’s supervenience base; (M*) stops requiring supervenience bases to be held fixed. Common causes of Ψ and Φᵢ now count as interventions, and MM is satisfiable, but (M*) then rules mutually manipulable Ψ and Φᵢ to be in causal feedback, which contradicts (NC).
• (M) (pp. 17–18).** X causes Y only if some intervention sets X at t and changes Y at a strictly later t′. Because (SC) makes macro and micro changes overlap in time, no constituent is ever a cause of its whole, and (NC) follows instead of being assumed. This is the first consistent embedding of MM.
• The evidential problem (pp. 18–20). Under (M**) every intervention that satisfies MM is a common cause of Ψ and Φᵢ. “Model 5(a), which features constitutive dependencies … and model 5(b), which does not … imply the same correlations under manipulations. They are empirically indistinguishable.” The asymmetry that matters: one can change the micro level without changing the macro level, but not the reverse (p. 20; also footnote 8: bottom-up interventions “may well exist”).
• (FH) and (CR) (p. 21). (FH): every (IV*)-intervention on Ψ is a common cause of Ψ and some Φᵢ ∈ V. (CR): V’s elements are constitutively relevant to Ψ iff MM and FH hold. FH has universal form, so “compliance with (CR) can, in principle, only be established inductively and is, thus, always prone to error” (p. 22).

Claims

id	claim	strength	support
C1	Given (Sup) and (NC), every cause of a mechanism’s macro behaviour is a common cause of it and of some constituent’s behaviour	strong	deduction from (Sup), (NC) and the interventionist definition of cause, pp. 11–13, 20
C2	MM with Woodward 2003 interventions entails that no constitutive relevance exists	strong	reductio, pp. 10–13; valid on its premises
C3	MM with Woodward 2011 interventions contradicts (NC)	strong	pp. 15–17
C4	With (M**), MM is consistent and entails (NC)	moderate	pp. 17–18; rests on (SC) making macro–micro changes strictly overlap, and on (M**), which the authors offer only for mechanisms (footnote 10)
C5	Mutual manipulability gives no first-order evidence for constitution	strong	Fig. 5, pp. 18–20
C6	Unresolvable fat-handedness of top-down interventions, found inductively, is abductive evidence for constitution	weak	a proposal; no case study (p. 22, “a project we have not even begun”)

Concepts

• Fat-handed intervention — “an intervention that influences its ultimate effect along two (or more) different causal paths” (footnote 11, after Scheines 2005).
• Specific variable — a variable for the behaviour of a particular entity, Ψ(S), Φᵢ(Xᵢ), after Spohn (p. 4).
• Surgical cause — a cause of Ψ that is not a common cause of Ψ and some constituent; FH says there are none (p. 20).

Connections

• Craver 2007 (LIT-832, unread here) and Craver 2007a are MM’s sources; Woodward 2003 and 2011 the interventionist background.
• Leuridan 2012 (LIT-1929) drops (NC); the paper answers him on p. 15.
• Baumgartner & Casini (LIT-1939, NOTE-1490) turn FH into the No De-Coupling account and drop the claim that a temporal clause rescues MM’s evidential role. Baumgartner, Casini & Krickel (LIT-1924, NOTE-1494) restrict fat-handed interventions instead.
• Romero 2015 (LIT-1933) reaches the fat-handedness of interventions on phenomena independently, by later citations.

Bearing on the record’s own argument

• CLAIM-522: left standing, and its “declared boundary” clause is sharpened. The paper’s problem is synchronic: whole and part change at the same time, so nothing can act on one without the other. CLAIM-522’s separating interventions are diachronic. Knocking out a production pathway in CASE-189 (removing c₂ → c₁) changes ċ₁ at once and c₁ only over the following interval, so by the paper’s own temporal criterion (M**) (p. 17) the pathway is a cause of the later capacities, not a constituent of them, and interventions on it can be surgical. The paper allows this explicitly: bottom-up interventions “may well exist” (footnote 8, p. 13). So the fat-handedness result does not reach CLAIM-522’s conclusion. What it does reach is the step CLAIM-522 already declines to take: whether the producer is part of the system. The paper takes parthood as given (SC) and refers its problems to Leuridan (p. 6). Intervention data do not supply a boundary here either, which is CLAIM-522’s “relative to a declared boundary” and CLAIM-194’s point. CLAIM-528 and ARG-050 set this out.
• CASE-189 holds a fat-handed dependence of exactly this kind. The coupling L(c) = Bᵀ diag(a₁, a₂)B with aᵢ = max{cᵢ − ½, 0} is a function of the capacities, changing whenever they do. Any intervention on c is an intervention on L, so “c maintains the coupling” and “the coupling is realized by c” cannot be separated by intervening on c. CLAIM-521’s comparison of coupling and maintenance is made on the model, which is what makes it possible. Same structure, not shown to be the paper’s part–whole relation: L is a parameter, not a whole.
• CLAIM-515’s third question, “Realized … shown by interventions that discriminate implementations”. If “realized” means that the capability consists in the mechanism, synchronically, this paper says intervention data cannot show it: the interventions that change both are common causes (C5), and only a pattern across many experiments supports it (C6). If it means that the capability is produced through the mechanism, diachronically, interventions can show it. CASE-163’s identification of a mechanism from probes within a catalog is of the second kind. The claim does not say which it means (CLAIM-528).
• QUESTION-284: an answer for one kind of constitution. QUESTION-284 asks “whether any evidence distinguishes ontological constitution from ordinary causal modification”. For part–whole constitution in mechanisms, this paper’s answer is: no single manipulation does (C5), because every manipulation of the whole is a causal modification of some part. A second-order pattern does (C6), and only inductively. That is CLAIM-150’s underdetermination with a way out, and the way out is abductive. The question’s “constitute” is about relations fixing the identity of relata, which is not this relation; the bearing is the form of the evidential problem, not shown to be the same problem.
• CLAIM-388 and CLAIM-386: the asymmetry the closure tests use. The paper’s asymmetry, that the micro level can change without the macro level changing but not the reverse (p. 20), is exactly what a test of dynamical closure uses. CLAIM-388’s sum is closed because changes in the distribution of activity among units that leave the sum fixed do not change the sum’s future; CLAIM-386’s closure defect QA(I − VQ) is the aggregate’s response to such within-block redistributions. Those are bottom-up interventions inside a fibre of the supervenience map, and they are the surgical ones. Closure of a collective variable is thus testable by the interventions this paper allows, which constitution is not.
• CLAIM-526: coexisting levels without inter-level causation. The paper’s (M**) makes macro and micro “two closely related, but still distinct” relata of a non-causal dependence (pp. 22–23), which is CLAIM-526’s coexistence. The levels differ, though: the paper’s are levels of parthood; CLAIM-526’s are levels of description fixed by closure. CLAIM-405 already says such criteria need not agree.
• TERM-088. The term’s “constitutive” sense (relational organization helps fix identity and boundary) is not this paper’s constitutive relevance (a part’s behaviour realizing a whole’s behaviour). A claim citing this debate should say it means the latter.

Limitations

• (NC) and (Sup) carry everything. Drop (NC) and the first reductio fails (Leuridan’s route). Make the base the organized activities rather than the behaviours and the argument for using (b) depends on Woodward’s difference-making view of causal structure (pp. 5–6).
• (M) is offered “only” for causation in mechanisms (footnote 10), and the strict temporal overlap of macro and micro changes is assumed, not argued. Where a part’s response takes time, (SC)‘s overlap is an idealization (CLAIM-480’s point; CLAIM-528 part 3).
• No case is analysed. The Mimosa example is illustrative; the authors say they have not tested whether practice fits (CR) (p. 22).
• FH is untestable as stated except inductively, and the paper gives no account of how large a variable-set expansion must be.

Open questions

• Does any real constitutive inference in neuroscience or biology proceed by searching for a surgical cause of the whole and failing to find one? A reconstruction of one would test FH.
• When the micro response lags the macro change, is the dependence causal by (M**)? If so, the line between constitution and causation moves with the time resolution of the description.