No universal distillation preserves all task-relevant structure
Type: kb/notes/types/structured-claim.md · Status: seedling · Tags: computational-model, context-engineering, learning-theory
Distillation works because it targets a specific consumer-task pair. The source persists because the distillate is lossy. This note makes that necessity explicit. Fix a state space X, a family of downstream queries Q, and a summary budget of b bits. A universal distillation for Q is a summary map sigma: X -> {0,1}^b plus a decoder delta such that for every state x and every downstream query q in Q, the answer q(x) can be recovered from sigma(x) and q alone. Such a distillation exists only if the summary space is large enough to separate all query-relevant distinctions in X. Equivalently: if Q induces more than 2^b distinct answer profiles over X, then no b-bit universal distillation exists. In particular, for sufficiently rich query families on growing states, no bounded-size summary can preserve all task-relevant structure.
Evidence
- The KB's own definition of distillation already says the target is always a specific consumer-task pair and that multiple distillations of the same source are normal. This theorem explains why that targeting is not just practical advice but a structural necessity.
- The right notion of "task-relevant structure" is not every detail of
x, but the partition of states induced byQ. Two states are equivalent if every query inQgives the same answer on both. Any exact summary only needs to preserve the equivalence class, not the raw state. - A summary with
bbits has at most2^bpossible outputs. So it can distinguish at most2^bequivalence classes or answer profiles. - Rich query families can induce exponentially many profiles. Canonical example:
X = {0,1}^N,Q = {q_i(x) = x_i : i in [N]}together with arbitrary downstream Boolean functions of the bits, or simplyQ = {q_S(x) = x_S : S subseteq [N]}. Then distinct states have distinct answer profiles, so exact universal summarization requiresOmega(N)bits.
Reasoning
For each state x in X, define its answer profile under Q as
A_Q(x) = (q(x))_(q in Q).
Two states are task-equivalent exactly when they have the same answer profile. An exact universal distillation needs only to preserve A_Q(x), but it must preserve that profile completely.
Suppose there are more than 2^b distinct answer profiles but the summary has only 2^b possible outputs. By the pigeonhole principle, there exist two states x != y such that
sigma(x) = sigma(y), butA_Q(x) != A_Q(y).
Since the profiles differ, there exists at least one query q in Q such that q(x) != q(y).
Now run the decoder on that query. Because sigma(x) = sigma(y), the decoder sees exactly the same inputs in both cases: the same summary and the same query. So it must produce the same output on both inputs:
delta(sigma(x), q) = delta(sigma(y), q).
But exactness requires that output to equal q(x) in one world and q(y) in the other, and those values differ. Contradiction.
Therefore any exact universal distillation for Q needs at least as many summary codes as there are distinct query-induced answer profiles. In bit terms, the minimum summary size is at least
log_2 |{A_Q(x) : x in X}|.
This yields the promised impossibility statement. If the number of query-induced profiles grows without bound with state size while the summary budget stays fixed, then eventually the summary space is too small and a universal distillation becomes impossible.
The practical consequence is that "compress once and discard the originals" is sound only for a restricted downstream query family known in advance. General-purpose memory does not have that property: future tasks may ask queries that separate states the summary collapsed. So a general memory system must either keep the originals, keep pointers back to them, or accept re-distillation and lossy failure on some future tasks.
Caveats
- The theorem is about exact worst-case universal preservation. Approximate summaries, probabilistic guarantees, or task distributions can still be useful in practice.
- The lower bound is only as strong as the downstream query family. If
Qis coarse, a very small summary may be enough. This is not an anti-distillation result; it is a formal statement that distillation must be targeted. - Keeping a bounded summary plus access to the original state does not violate the theorem. The impossibility is about the summary being sufficient on its own after the originals are discarded.
- Interactive recovery changes the model. A system that can go back to storage, reopen sources, or ask additional bounded calls is no longer relying on one fixed summary artifact to answer every future query.
Relevant Notes:
- distillation — foundation: this theorem formalizes why distillation must be targeted to a consumer-task pair rather than treated as lossless compression
- evolving understanding needs re-distillation not composition — consequence: when downstream needs change, the right response is re-distillation, not faith that one old summary preserved every future distinction
- bounded-context orchestration model — context: compaction of symbolic state is a summary map, so the theorem identifies the structural limit of "summarize K and keep going"
- LLM-mediated schedulers are a degraded variant of the clean model — instance: conversation compaction is precisely the kind of bounded summary this theorem says cannot be universally sufficient
- exact retrieval over semantically opaque items requires linear inspection — parallel lower bound: one rules out universal compression, the other rules out universal pruning without semantic inspection