Contribution
It asks whether safety constraints permit the informative experiment that adaptive control needs, rather than how fast uncertainty shrinks once they do. Most finite-time analyses assume persistent excitation of the uniformly safe closed loop, under which “feasibility is already settled; only the rate remains” (abstract). The paper identifies the object that decides feasibility, the commitment rule, and reduces safe attainability of the oracle to three quantities it induces (Theorem 1). It then proves linear regret in two constrained linear–quadratic examples, gives a list-size converse for many alternatives (Proposition 1), proves recovery under a repeatable-experiment condition (Theorem 2), and gives SDP upper certificates for the information quantity (§6).
Key insight
Evidence must precede commitment, and the margin that commitment spends can also be what generates the evidence. “Because the commitment decision is made from (H_{t−1}, U_t), the observation produced by U_τ arrives too late to justify that action” (p. 2). So the full trajectory is “the wrong statistical experiment”; the right one is the record stopped at commitment. If the actions that tell two models apart consume a safety margin that must be preserved under the alternative, the information available before commitment is capped by the margin, however long the horizon.
Assumptions
- Uniform chance safety (Def. 1, eq. 7): the same causal policy keeps P(violation) ≤ η_T under every θ in the initial class Θ₀; typically η_T = T⁻². Performance is measured against the oracle that knows θ and is safe for θ alone (eq. 9).
- Commitment (Def. 3): for an ordered pair (θ, θ′), a predictable set of history–action pairs; in deterministic systems, the first action after which no hard-safe continuation remains under θ′. τ is a stopping time for F_t = σ(H_{t−1}, U_t).
- Lemma 1 needs absolute continuity of the observation kernels; the theorem itself does not.
- The lower bound needs no Markov property, linearity or Gaussian noise (Remark 5). The examples are deterministic or linear–Gaussian; the recovery theorem treats a scalar parameter with a known conservative direction.
Key results
- Stopped KL identity (Lemma 1). D(P^pre_θ ∥ P^pre_θ′) = E_θ[Σ_t 1{t < τ} kl_t], where kl_t is the one-step KL of the observation kernels. The policy kernels cancel, including the one that generates the committing action.
- Theorem 1 (binary limit, p. 10). With α_T = sup P_θ′(commit) over uniformly safe policies, G_T(θ, θ′; q) the least target regret among safe policies that fail to commit with probability ≥ q, and I_pre = sup stopped KL: R_T(π; θ) ≥ G_T(θ, θ′; ψ_{α_T}(I_pre)) ≥ G_T(θ, θ′; [½e^{−I_pre} − α_T]₊). With a linear profile G_T(q) ≥ [G_T q − ζ_T]₊ this is R_T ≥ G_T(½e^{−I_pre} − α_T − ζ_T/G_T). If β_pre = sup_T I_pre < ∞, α_T → 0 and ζ_T = o(G_T), every uniformly safe policy has worst-case regret at least ½e^{−β_pre} G_T asymptotically (eq. 37). Proof: data processing on the event “no commitment” and Bretagnolle–Huber.
- The evidence threshold. The explicit bound stays positive until I_pre ≥ log(1/(2α_T)), which is 2 log T − log 2 at α_T = T⁻² (p. 11).
- Corollary 2. Regret r_T = o(G_T) requires unbounded pre-commitment information whenever α_T → 0.
- Finite headroom (pp. 3, 7). Margin d_{t+1} = d_t − θu_t², sensor y_t = θu_t + ξ_t. Before leaving the worn device’s budget Σu² ≤ d₀/θ_d, so I_pre ≤ (θ_d − θ_b)²d₀/(2σ²θ_d), independent of T. For a healthy target θ_b = 0 the gap is Θ(T) and regret is linear; for 0 < θ_b < θ_d it is Θ(√T). On a continuum Θ₀ = [0, θ̄] the worst alternative approaches the boundary target at rate T^{−1/3} and R_T ≥ (UT/2)(1 − o(1)) (eqs. 44–46).
- Corollary 4 (linear regulation, p. 18). Two variants, one-sided linear margin and signed quadratic margin, both give R_T ≥ c_i T. A cross-loading vanishing as T^{−1/2} still blocks the whole linear advantage (eq. 68).
- Proposition 1 (safe lists). To return a list of size ≤ L containing the true one of M alternatives with error ε, I(J; S) ≥ (1 − ε)log(M/L) − h₂(ε). With finite sensing, partial adaptation through a certified list is still possible: at N = 36 in §7.3 the list policy recovers 27.6% of the robust-to-oracle gap.
- Theorem 2 (recovery, p. 22). If a conservative experiment is safe, repeatable from a common state, informative (V_t grows) and its value loss is bounded by a modulus ω of the parameter error, a lower-confidence rule is chance-δ safe with regret ≤ r_max + Σ E ω(2b(V_{t−1}, δ; ν)) + r_max Tδ. Braking example: O(σ√(D/θ)·√(T log T)) (Corollary 7).
- Certainty equivalence fails at an active boundary (§5.6.1). Plugging in the sample mean makes each braking episode unsafe with probability exactly ½, for every n; more data shrink the overshoot, not its probability.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Bounded information before commitment forces a fixed fraction ½e^{−β} of the noncommitment gap as regret for every uniformly safe policy | strong | Theorem 1, data processing and Bretagnolle–Huber; relaxation checked numerically |
| C2 | In a finite-headroom device and in constrained linear regulation the information is bounded and regret is linear | strong | (22), Corollary 4; the headroom bound and the boundary optimization checked |
| C3 | Finite information can still certify a smaller candidate list and partial adaptation | strong / moderate | Proposition 1 (Fano-type converse); Monte Carlo of one elimination rule |
| C4 | Repeatable, informative, safe experiments allow oracle recovery at O(√(T log T)) | strong, narrow | Theorem 2, proof read; scalar ordered parameter, common reset |
| C5 | Changing the estimator cannot remove a bounded-information obstruction; the experiment, the sensor or the degree of specialization must change (p. 36) | moderate | follows from C1; §7.1–7.2 analytic illustrations (replenishment, level sensor) |
Method
The lower bound is a recipe (§3.6): define a predictable commitment event; bound α_T under the alternative; bound the noncommitment regret profile G_T; bound the stopped KL. For deterministic linear–Gaussian classes the last step is a QCQP whose SDP dual gives upper certificates (§6).
Concepts
- Commitment — “the first action that forecloses” a safe continuation under an alternative (abstract).
- Precommitment information — “the KL divergence between learner-visible laws stopped before commitment” (abstract).
- Guarantee or exclude — the two options of a controller with uncertainty unresolved: stay where every plausible model is safe, or act as if an alternative were no longer in play; “The exclusion is statistical, not literal” (p. 2).
- Oracle regret — against the safe oracle that knows the realized model; the max–min criterion does not distinguish guaranteeing from excluding (p. 2).
Connections
- Persistent-excitation analyses of constrained LQR (Dean et al., Li et al., Hutchinson et al.) are placed “in a persistent precommitment-information regime” (p. 35).
- Dual control and controlled sensing (Chernoff-style sequential design) are the nearest frameworks; the paper adds the decision-time constraint and endogenous admissibility.
- Recursive feasibility, safety filters and returnability are the constrained-control tools; noncommitment is “exactly preservation of hard-safe recursive feasibility under that alternative” (p. 35). In the record that is viability (THEORY-134): a commitment is the first action that leaves the alternative’s viability kernel.
- Budgeted cumulative-violation constraints fall outside the theorem (Remark 4).
Bearing on the record
- The registration’s description is accurate in substance. Two precisions: the “fixed fraction” is ½e^{−β_pre} of the noncommitment gap G_T, net of α_T and ζ_T, against the safe oracle; and the paper also proves the positive side (Theorem 2) and partial adaptation by lists, which the summary omits.
- CLAIM-312 and TERM-085. A formal companion to both. CLAIM-312 says failure by the deadline establishes unavailability. Saligrama adds an information-theoretic floor on the evidence needed before the deadline: about log(1/(2α)) nats to act on an exclusion at risk α. The evidence a corrective act itself generates does not count towards that floor. For TERM-085’s correctability, a corrective objective that requires committing before enough evidence can be gathered has correctability bounded away from the oracle, whatever the update rule. That is A20’s use, and it holds.
- CLAIM-511 and CASE-128. The “guarantee or exclude” choice is a clean case of the disclaimer’s point. Staying correctable (preserving a safe continuation under every live alternative) has a price, ½e^{−β} of the gap, that is fixed by the information geometry, not by preference.
- QUESTION-285 (inquiry) and CASE-186. The sensor ablation (§7.2) is a crisp instance: the same safe prefix gives bounded evidence with a rate sensor and linearly growing evidence with a level sensor that records a persistent imprint. Whether a record of the past persists is what decides whether inquiry can continue.
- No instruction for ML practice beyond the recipe for control analysis; no anthology flag.
Limitations
- The lower bound is a sufficient certificate; failure of the certificate does not prove recovery (p. 36).
- Pairwise: a large pairwise profile can hide that no single safe policy resolves all alternatives (Remark 6, p. 36).
- Recovery is proved only for a scalar ordered parameter with a common reset; vector parameters and non-resetting dynamics are open.
- The SDP certificate omits process noise; the case of unbounded but sublinear pre-commitment information is unresolved (p. 36).
- Numerical sections are illustrations; safety is analytic throughout.
Open questions
- Is there a converse matching Theorem 2 when information grows sublinearly?
- For many alternatives, what replaces the pairwise profile when one policy must inform about all of them at once?