At the urging of my colleague Bill Idsardi, I went looking recently for examples of idioms that obligatorily involve syntactic movement as a necessary part of the idiom. This search landed me, among other places, in Nunberg, Sag & Wasow 1994 (henceforth, NSW94), and specifically, exx. (47‑48) on p. 516. There, NSW94 puts forth what it takes to be a collection of idioms (more on this shortly) where the availability of the idiomatic reading depends on the application of some transformation (more on this shortly), and is thus not available if everything stays in its base position. To me, lots of the examples NSW94 gives look a lot more like entirely compositional expressions with some cultural significance attached to them (such as Rome wasn’t built in a day) than idioms per se. Several others may have been seen as the result of “transformations” at some point in the development of syntactic theory, but are no longer seen as such. An example is break a leg, which is listed in NSW94 as involving the “imperative deletion” transformation (presumably, deletion of the subject…?). Let’s set aside the fact that this expression isn’t actually restricted to the imperative at all (cf. I told them to break a leg), and suppose that the description were true. On the current view of imperative syntax, there is likely functional structure in imperative clauses that is simply missing in their non-imperative counterpart (see, e.g., Zanuttini’s 2008 arguments for JussiveP in imperatives). So it is entirely within the realm of possibility that there would be an idiom that crucially depends on the presence of that imperative-specific functional material for the idiomatic reading to arise. It is beyond clear that the availability of an idiomatic reading can depend on functional material in this manner: kick the bucket vs. #kick a/some/his bucket, dying to VP vs. #died / will die to VP, and so forth (on the latter, see Harwood 2015).
However, while these confounds plague the majority of NSW94’s examples, a few, it seems to me, remain standing. I’ll focus in on one of them here:
(1) hard to take
The meaning of this expression is something like “difficult to bear.” Given that one of the meanings of the polysemous hard is a (near-)synonym of difficult, we might wonder whether take is simply polysemous as well, with one of its meanings being a (near-)synonym of the relevant meaning of bear. Call this a ‘distributed synonymy’ analysis of the meaning of (1). I think this account is wrong, however. Consider the following:
a. This situation is difficult to bear.
b. It is difficult to bear this situation.
a. This situation is hard to take.
b. ??It is hard to take this situation.
The ‘distributed synonymy’ analysis leaves (3b) unexplained, especially in contrast to (2b). Note that this is not about the propensity of take vs. bear to appear with or without an object, as even the (a) examples involve movement of a phrase originally base-generated as the complement of the embedded verb. That is: all four of the examples in (2‑3) involve the embedded verb merging with a direct object.
It therefore appears that what we have in (1) is a bona fide movement-dependent idiom. (In this case, an idiom involving the tough-adjective hard that crucially depends on tough-movement having taken place.)
How could this be accounted for? Well, one obvious approach that suggests itself in light of the earlier discussion of imperatives is that tough-adjectives co-occur with different functional material in those sentences where tough-movement occurs (e.g. (2b)) vs. those where it does not (e.g. (2a)). We could imagine there is some FP occurring close to the adjective that bears a movement-triggering feature in the (a) examples but not in the (b) examples. The availability of the idiomatic reading of hard to take may then depend on it co-occurring with the movement-inducing variant of the F head, and be unavailable otherwise.
But this all leads me to a more general question, and one that was actually at the heart of the discussion Bill and I were originally having. It goes as follows. Suppose idioms are mappings from sets of syntactic terminals to individual units of meaning; or suppose they are conspiracies of mutually-conditioned allosemies of individual terminals (Harley 2014). Either way, the conditioning of these special readings is structural. (Kick the bucket doesn’t literally depend on the linear adjacency of the terminals in question, cf. kicked the bucket.) But a structure built by movement is also structure. Why can’t the structurally-stated conditioning environment for the idiomatic reading involve movement? Note that this question is especially pressing in light of the current view that unifies structure-building (External Merge) and movement (Internal Merge) under a single formal operation. And it is extra-especially-pressing if one adopts the view of movement as creating multidominant structures (the “Remerge” theory of movement). That’s because it is entirely unclear why the conditioning environments for idiomatic readings (whether the mapping is wholesale or through a conspiracy of allosemies) cannot make reference to the kind of Directed Acyclic Graph (DAG) that a multidominance syntax builds.
The hard to take idiom in (1) is plausibly, I think, an idiom that depends on movement. But interestingly, it doesn’t depend on movement of one of the terms of the idiom. (I.e., the DP in DP is hard to take is an “open slot” in the idiom.) Are there idioms where the availability of the idiomatic reading depends on movement of one of the terms of the idiom? And if not, how exactly would one block this possibility, theoretically speaking?
Either way, it would be great to know the answer. In a sense, it would be especially interesting if it turned out that such idioms were categorically unavailable, since the question at the end of the previous paragraph would apply in full force; and it strikes me as a potentially interesting one.
As an aside, the discussion above made a handful of assumptions that seem to contradict the prevailing wisdom, in some circles, concerning idioms. I’d therefore like to flag these assumptions and briefly defend them. The first is that idioms can include, and depend on, functional material. This one was already defended above.
A related claim one encounters in the literature sometimes is that while idioms can include functional material, the interpretation of that material is the same in the idiom as it would be in a non-idiomatic expression. Thus, the thinking goes, it is no accident that kick the bucket is telic even on its idiomatic reading, because it has the syntax of a telic verb phrase. I believe this generalization also fails to hold. Shoot the breeze (meaning “talk at length”) is atelic on its idiomatic reading, but insofar as kick the bucket is a ‘telic verb phrase’, so is shoot the breeze. Anecdotally, some have the intuition that breeze feels like a mass-noun; whatever we want to make of that intuition, we know that properties of the denotatum (like homogeneity or countability) are neither necessary nor sufficient for mass-noun-ness (cf. furniture). And, more to the point, a breeze simply doesn’t require the implicit quantization that things like a beer or a water (each meaning one standard-sized serving of the relevant liquid) require; cf. also *an air.Thanks to Roger Schwarzschild for helpful discussion. I therefore maintain that shoot the breeze would be predicted to be just as telic as kick the bucket (contrary to fact).
Next: idioms are not bounded by phases.This stands in opposition to the assumptions of the very same Harwood 2015 paper cited above, by the way. To count X’s chickens before they hatch has a whole darn finite clause inside it. And even simpler examples like to yank X’s leg (meaning “to jokingly mislead or deceive X”) contains the DP [X‘s leg], which is almost certainly a phase.
Surely, the generalizations just surveyed are mostly true, but they are not combinatorially true, which is how some of the literature sometimes presents them. Especially when it comes to the phase issue, I think there’s likely a learning explanation for why the phase-boundedness generalization is probabilistically true, and I suspect that the grammar (in the narrow sense) has nothing to do with it.