Craver (2008) sharply distinguishes between two traditions of understanding scientific explanation: reductive explanation and systems explanation. According to Craver, the first tradition accepts a version of the covering law model of explanation (Hempel 1965) whereby one explains regularities at a given level of organization by showing how these regularities (the laws describing events and their relations) can be derived from theories holding at lower levels. Put differently, one explains a phenomenon of interest by showing how it is to be expected based on the laws governing activity at lower levels of organization. This tradition is reductive because when such explanations are successful, one can strictly speaking do without the higher-level laws. However convenient they may be for understanding or predicting higher-level phenomena, the higher-level laws do not add, capture, or explain any facts that are not already contained in the lower-level laws. The lower-level laws are scientifically sufficient.
In contrast, in the systems tradition, a phenomenon of interest ψ exhibited by a system S is explained by identifying a set of component parts {X} and showing how they are organized such that S ψs. A systems explanation is similar to reductive explanation in that it too relies on the identification of levels of organization, since it requires identifying the parts of the system S, but, as I note below, it does not aim thereby at the reduction or explanatory absorption of one level by another. Craver & Bechtel write:
In levels of mechanisms, an item X is at a lower level than an item S if and only if X is a component in the mechanism for some activity ψ of S. X is a component in a mechanism if and only if it is one of the entities or activities organized such that S ψ’s. For that is what mechanisms are: they are entities and activities organized such that they exhibit a phenomenon. Scientists discover lower levels by decomposing the behavior of a mechanism into the behaviors of its component parts, decomposing the behaviors of the parts into the behaviors of their parts, and so on. (2007, pp. 548–549)[1]
As already noted, S is the system that ψs, or that exhibits phenomenon. It is, for instance, the car (S) that accelerates (ψ), and to explain car acceleration will require identifying the components {X} that matter to S ψ-ing. To identify these components and their organization is to explicate the mechanism M that accounts for S ψ-ing. The target of mechanistic explanations of this sort is ψ: “mechanistic explanations are framed by the explanandum phenomenon” (Craver 2008, p. 121) and “[t]he explanandum of a mechanistic explanation is a phenomenon, typically some behavior of a mechanism as a whole” (Craver 2008, p. 139).
In mechanistic explanation, a given X is a component of the mechanism M if and only if it is one of the entities organized such that S exhibits some phenomenon ψ. So the engine, the accelerator, and the gas tank, but not the mud-flaps or the windshield wipers are components of M that explain the car accelerating, even though these are all parts of the car S. In an ideal explanation, the mechanism defined by the parts {X} will contain all and only the components relevant to S ψ-ing (see Craver 2008 for a discussion of constitutive relevance in this context). To identify the parts of M is thus to specify both a hierarchical and a functional relationship between M and its parts, and between M and S.
But although mechanistic explanation involves essential reference to hierarchical relationships between levels of organization, it is not thereby a species of reductive explanation because in a successful systems explanation nothing is rendered inessential or redundant. The phenomenon ψ is neither derived nor derivable from laws governing the parts of M; rather, the parts {X} and their relationships simply are M, and together explain why S ψs. The explanatory relationship is not rational derivation, but functional composition: M is physically and functionally constituted by its parts, and S ψs in virtue of that constitution.
Mechanistic explanations are constitutive or componential explanations: they explain the behavior of the mechanism as a whole in terms of the organized activities and interactions of its components. Components are the entities in a mechanism—what are commonly called ‘parts’. (Craver 2008, p. 128)[2]
Given all this we can add one more criterion for a given X being a part of the mechanism M: each X must be not just a functional but also a spatial sub-part of M. As a component of M, X will be at a lower level than M, and smaller than M: “[b]ecause mechanisms are collections of components and their activities, no component can be larger than the mechanism as a whole, and so levels of mechanisms are ordered by size” (Craver & Bechtel 2007, pp. 549–550). Craver and Bechtel conclude: “[m]ost fundamentally, levels of mechanisms are a species of compositional, or part-whole relations” (Craver & Bechtel 2007, p. 550). In the overall framework developed by Craver and Bechtel, functional levels and spatial levels generally align.
Thus, although componential mechanistic explanations are not reductive, they generally are what one would call “bottom-up”, or perhaps better in this context, “level-restricted”: one explains the phenomenon ψ in S by reference to entities and relations at a lower level of organization, but never the reverse. In componential explanations of this sort, the intrinsic properties of and interactions between the mechanism’s components account for a system’s actions (where “intrinsic” means that such properties—such as the charge of an ion—are either basic to the entity or accounted for by reference to entities and properties at a still lower level of organization). Good mechanistic explanations on this view will not include references to unanalyzed properties of the whole S or M, its “shape” or overall organization, as the relations between the components {X} at the lower level will already account for (in fact constitute) these.
This account of mechanistic explanation seems to me a clear and, indeed, compelling model of one kind of explanatory practice in the neurosciences. To satisfy the norms of mechanistic explanation, one must:
Identify the phenomenon of interest ψ
Identify the system S that ψs
Identify the relevant spatial sub-parts {X} of M (and their relevant intrinsic properties)
Describe how the parts {X} are organized such that S ψs
At least prima facie, a number of instances of successful (albeit incomplete) explanatory models in the neurosciences appear to neatly fit this description. Craver (2008) extensively discusses the mechanistic model of the action potential. Briefly, following the steps above:
The phenomenon ψ is the action potential, which consists of the rapid depolarization of neural cells from a resting membrane potential of approximately –70mV toward (and in many cases significantly exceeding) 0mV; an equally rapid repolarization; a period of hyperpolarization, where the cell overshoots the normal resting potential; and a gradual return to the resting equilibrium (note that as even this simplified sketch illustrates, ψ will often be in and of itself complex, with many aspects that any adequate model must capture).
The system S that ψs is the neuron.
The parts in virtue of which S ψs include elements of the cell and its surrounding ionic milieu: positively charged K+ and Na+ ions; gated, ion-specific membrane channels; and the Na+/K+ pump.
Finally, the organization that explains ψ includes the following: The resting potential is in fact an equilibrium between two opposing forces: a chemical concentration gradient that pushes Na+ into the cell and K+ out of it, and an electrical gradient that pushes K+ into the cell, each maintained by the selective permeability of the cell to Na+ and K+. Na+ channels change their conformation in response to current flow (they are voltage-gated) such that they open to allow Na+ to flow into the cell. As Na+ flows into the cell this reduces the electrostatic pressure on K+, and opens voltage-gated K+ channels, allowing K+ to flow out of the cell. The net effect is to push the cell initially toward the electrochemical balance point for Na+, which is about +55mV. However, as the membrane potential drops, the Na+ channels close, thus slowing and eventually stopping the depolarization. The diffusion of K+ out of the cell combines with the activity of the Na+/K+ pump to repolarize the cell, which however overshoots the resting potential due to the fact that the K+ channels close later than the Na+ channels, thus allowing K+ to diffuse out of the cell for an extra millisecond or so during which the cell is hyper-polarized.
Obviously, this remains a sketch (see Craver 2008) or any basic neuroscience textbook for more detail), but it illustrates the main elements of a mechanistic explanation. The intrinsic properties, actions, and interactions of M’s spatial sub-parts together comprise the mechanism that allows S to ψ and thus explain how S ψs. One can likewise plausibly sketch the mechanisms that account for spatial long-term memory (e.g., the ability of an animal to return to some location in its environment) in terms of long-term potentiation of synapses in the hippocampus (Craver 2008), although it is worth noting that a more complete account of the functions of hippocampus will have some of the features I describe in 2 and 3 (Buckner 2010; Anderson 2015). Still, the fact that some explanations in neuroscience are like this is not under significant dispute.
But this brings us to the question of why I have distinguished M and S in my treatment. Because Craver (2008) does not formally distinguish these, he is never led to ask what the precise relationship is (or could be) between M and S (and between their respective parts). In fact, for Craver the symbol S usually (but not always) refers to what I have been calling M, and he frames his analysis of mechanistic composition entirely in terms of ψ and its mechanism. When he does mention the larger system it is generally to emphasize the fact that not every part of a system S is relevant to the mechanism in virtue of which it ψs. So what might the committed mechanist say about the relationships between S, M and {X}? One possibility is: all the parts {X} of M will be on a lower level than S. That would be in keeping with the level-restricted character of the framework, and its characteristic alignment between spatial and functional levels. It is certainly a feature of all the examples discussed in its support, including the model of the action potential outlined above. A slightly stronger possibility would be: all the parts {X} of M will be spatial sub-parts of S. I don’t think anyone would or should endorse this stronger condition, but seeing why will be instructive, and will lead us to the reasons to reject the weaker formulation as well.[3]
The immediate trouble with the stronger formulation is that it collides with a fact noted by Craver (2008), but not otherwise discussed: the mechanism that accounts for S ψ-ing may contain parts that are extrinsic to S (although not to M). For instance, in the mechanism for the action potential, the Na+ and K+ ions that are clearly part of M are (at least sometimes) extrinsic to S; and in embodied accounts of some cognitive processes like mathematics, the mechanism that accounts for a person (P) multiplying (ψm-ing) contains parts that are always extrinsic to P, such as pencil and paper (Clark 1997; see also this collection). These entities would arguably not be components of the systems that ψ, although they would be components of the mechanisms in virtue of which they ψ. At the very least, this suggests there are some details yet to be worked out about the necessary physical relationships between M and S that implement the hierarchical and functional relationships in virtue of which M can account for S ψ-ing. There will be (presumably rare) cases in which M and S are identical; cases such as the accelerating car where M contains only parts of S; and cases such as the action potential where M and S cross-cut one another, sharing some but not all of their parts.[4] There may also turn out to be cases in which they share no parts, perhaps because the parts of M and the parts of S are individuated by different criteria, or because S’s ability to ψ is imposed by or inherited from an entirely extrinsic mechanism (indeed I’ll discuss a potential instance of this class of cases later in the paper).
But distinguishing M and S in this way also allows one to ask whether all the parts of M need to be at a lower level than S. If not every X needs to be a spatial sub-part of S, then there is little reason to suppose that each X needs to be on a lower level than S, either. Indeed, I claim that in fact for some systems S the mechanism M will contain items that are neither intrinsic to nor at a lower level than S. For instance, I often use other people to help me remember things, in the easiest case by asking them to remind me at some future time. In such a case, this other individual is arguably part of the mechanism responsible for my remembering, but is certainly not for that reason on a lower ontological level than I am, qua remembering system. Moreover, as I will argue when looking at the case discussed below, some relevant parts of M (and certainly M itself) are at a higher organizational level than S. Now of course, Craver & Bechtel define the concept of lower level in terms of being a part of the mechanism: “an item X is at a lower level than an item S if and only if X is a component in the mechanism for some activity ψ of S” (2007, p. 548). I agree that this holds for the constitutive relationship between mechanisms and their parts. But it only holds for all systems S if we assume that all the parts of M are parts of S, and we have seen that this is not always the case. Thus although I think that Craver correctly analyzes the relationship between mechanisms and their parts in terms of constitution, I argue that the more capacious notion of enabling constraint better captures the relationship between mechanisms and the systems whose activities they enable.
In any case, with this as background, I now turn to the case of the SAC. In 2, I describe what we know about how the mechanisms in virtue of which the cell operates, and in 3 I discuss the implications of this case for componential mechanistic explanation.