Introduction
The object, the motivation, and the question
The object. A deterministic machine running on an input x x x produces a path of configurations C 0 x , C 1 x , C 2 x , … C^x_0, C^x_1, C^x_2, \ldots C 0 x , C 1 x , C 2 x , … , one per step. We do not follow this path step by step. Instead, an observer looks at it only at indices chosen by a random clock. At each tick the clock may jump forward, step back, or stay where it is, and the observer sees the configuration at the clock’s current index (Figure 1 ). Once the machine accepts, its accepting configuration repeats forever, so the input is recognised as soon as the clock reaches or exceeds the first accepting index.
Figure 1. A tick updates the clock index X t X_t X t and supplies the observation C X t x C^x_{X_t} C X t x . Crossing means X t ≥ A X_t \ge A X t ≥ A at some observed tick; persistent acceptance makes that observation sufficient.
Why skip steps? Computations are often reused rather than carried out from scratch. A program can store a computed value and look it up when it is needed again [1 ] . Studies of chess expertise likewise illustrate the role of learned perceptual patterns and their recall [2 ] . Language models provide another contrast: a prompt may elicit written intermediate steps, or an answer directly [3 ] . These examples motivate asking what is gained when access to a later state replaces inspection of the intervening states. The extent of a shortcut may vary; we represent this variation by a random clock, whose jump distribution may also depend on the current position. We idealise access: a shortcut lands on the true configuration. Uncertainty concerns how far one jumps, not whether the result is correct. The examples motivate the abstraction; they do not supply its configuration-access mechanism.
Why look back? A saved earlier state can be revisited, making another route forward available. In our clocks a backward step returns to an earlier index; it does not reverse the deterministic computation. It can return the observer to a position from which a long jump starts. Backtracking in a reasoning trace is an informal analogue. In the constructions of § 6 , large forward jumps are available only at specified gate positions. A forward jump can overshoot the next gate, while a step back by one can recover it. Within the prescribed families, these returns give crossing rates that no permitted forward-only clock achieves. For independent proposals, deleting backward jumps never reduces reach (Theorem 6.4).
Why count looks? Classical running time charges each executed transition. To measure access along a fixed computation path, the jumpy observation model instead charges one unit per observed configuration, and nothing for producing configurations, sampling the clock, or writing indices. This separates the logical length of a computation from the number of observations needed to recognise its outcome. Within this abstraction, threshold-crossing probabilities determine the observation budget. The resulting classes describe an observation resource, without asserting an ordinary machine speedup or a physical implementation.
The question. The clock is fixed in advance, independently of the input and the machine; a position-dependent clock uses the current index, not the contents of the configuration. This restriction matters: an externally prescribed schedule X t = 2 t X_t = 2^t X t = 2 t reaches index 2 p ( n ) 2^{p(n)} 2 p ( n ) after p ( n ) p(n) p ( n ) observations, placing the access resource in the schedule itself. For a fixed random clock, the probability of reaching a threshold within q q q ticks determines the languages recognisable at confidence α \alpha α (Theorem 2.4). We ask which jump laws and position-dependent kernels let polynomially many observations recognise which languages, and when backward access enlarges this power beyond every comparable forward-only clock in a specified family.
Background. Didehvar developed the programme of fuzzy-time computation to reconsider the role of uncertain time in computation [4 , 5 ] . It motivates asking how access to successive configurations affects what an observer can recognise within a given budget. The present paper is a further attempt to formalise these ideas. As a first step in this formulation, we study one perspective: stochastic access to a fixed deterministic computation path and the resulting observation complexity. A random clock selects indices by moving forward, backward, or remaining in place. This perspective separates the computation’s logical length from the observations needed to reach its accepting tail. Clocks with independent proposals are held random walks with one fixed law μ \mu μ , called fuzzy clocks in Theorem 2.2; we also consider position-dependent kernels. For a fixed clock, “fuzzy” concerns probabilistic uncertainty in the observed time index, rather than fuzzy-set membership. Probability specifies that clock’s motion. Possibility cuts [6 ] express incomplete specification by selecting allowed clock parameters; Appendix § 14 studies observation budgets common to all the permitted parameters.
A gate that a forward jump misses
The doubling gates of Theorem 6.12 illustrate the accessibility question. At a gate g = 2 k g = 2^k g = 2 k , the clock moves to 2 g + 1 2g + 1 2 g + 1 , just beyond the next gate 2 g 2g 2 g . A return by one reaches that gate, where the next large jump is available. A forward-only clock cannot revisit the gate just passed. If its forward displacement away from gates is bounded by a fixed integer w w w , accessing another gate requires travel using those bounded steps. The proof of Theorem 6.13 charges each gate’s gain to this intervening travel and bounds the reach of every forward-only member by a linear function of the tick budget. Length-one returns instead allow logarithmic crossing budgets. Squaring gates give the corresponding polynomial forward-only bound and doubly logarithmic budget (Theorem 6.18 ).
This comparison concerns all forward-only kernels within the prescribed gate families. The parity clock makes a weaker comparison: deletion changes its class from exponential to polynomial time, but replacing returns with forward re-entry also restores exponential time (Theorem 6.7). Independent-proposal clocks satisfy a deletion obstruction: suppressing negative proposals never decreases their reach (Theorem 6.4).
Main results
For a clock started at zero and a fixed confidence 0 < α < 1 0 < \alpha< 1 0 < α < 1 , let R α ( q ) R_{\alpha}(q) R α ( q ) be the greatest index crossed by tick q q q with probability at least α \alpha α . Persistent acceptance gives the reachability equivalence of Theorem 2.4: a recogniser belongs to the class exactly when its accepting times are bounded by R α ( Q ( ∣ x ∣ ) ) R_{\alpha}(Q(|x|)) R α ( Q ( ∣ x ∣ )) for some polynomial budget Q Q Q . Sequential input access makes small radii restrictive, while large radii permit longer computations to be observed.
The two principal construction results concern the order of classes below P and the necessity of backward access within gated families. They are followed by the deletion comparison and the iid tail hierarchy. The zero-drift example below is separate from the negative-drift antichain; no ordering between them is established.
Theorem A (Below P : locality and an antichain; Theorems 4.1 , 4.6 and 4.8 ) Negative drift with positive forward probability and all positive moments finite gives exactly the reverse-definite languages. One computable zero-drift law has a class strictly between the reverse-definite languages and P , and computable negative-drift laws give infinitely many pairwise incomparable such classes, at every pair of confidences.
For the fixed-return gated clock of § 6.3 , β ∈ ( 0 , 1 ] \beta\in(0,1] β ∈ ( 0 , 1 ] is the probability of a one-step backward return at a landing index.
Theorem B (Backward moves are necessary, and may be sparse; Theorems 6.13, 6.15 , 6.16 and 6.18 ) In a family whose only large forward moves are doublings at powers of two, every forward-only clock has class within P , whereas steps back by one, confined to a set of indices of density zero and possibly with vanishing probability, give EXP . The fixed-return gated clock has crossing budget ( 1 + o ( 1 ) ) ( 1 + β − 1 ) log 2 T (1 + o(1))(1 + \beta^{-1}) \log_2 T ( 1 + o ( 1 )) ( 1 + β − 1 ) log 2 T at every confidence. With squaring gates, backward indices of counting order log log N \log\log N log log N give 2EXP while forward-only clocks stay within P .
For the return-enabled parity clock of § 6.2 , β ∈ ( 0 , 1 / 2 ] \beta\in(0,1/2] β ∈ ( 0 , 1/2 ] is the backward-return probability at a positive even index. Write ℓ α = − ln ( 1 − α ) \ell_{\alpha} = -\ln(1 - \alpha) ℓ α = − ln ( 1 − α ) for 0 < α < 1 0 < \alpha< 1 0 < α < 1 , where ln \ln ln is the natural logarithm.
Theorem C (When backward moves help; Theorems 6.3 , 6.6 and 6.11 ) If the forward-displacement law is stochastically nondecreasing in position, deleting backward moves never lowers the class. For a parity-dependent clock, a step back by one reopens a heavy-sampling phase: the class rises from P to EXP and the crossing budget falls from ( 1 + o ( 1 ) ) T (1 + o(1))T ( 1 + o ( 1 )) T to ( 1 + o ( 1 ) ) ( 1 + β − 1 ) ℓ α log 2 T (1 + o(1))(1 + \beta^{-1})\ell_{\alpha}\log_2 T ( 1 + o ( 1 )) ( 1 + β − 1 ) ℓ α log 2 T , where ℓ α = − ln ( 1 − α ) \ell_{\alpha} = -\ln(1 - \alpha) ℓ α = − ln ( 1 − α ) .
Theorem D (Above P: the tail determines the class; Theorems 5.1, 5.5, 5.7 and 5.8) *For independent jumps the reach R α ( q ) R_\alpha(q) R α ( q ) lies between the α \alpha α -quantile of the largest of q q q proposals and q q q times that quantile. Consequently one fixed computable law has class exactly EXP ; for each k ≥ 1 k \ge1 k ≥ 1 one fixed computable law has class exactly Q P k = D T I M E ( 2 O ( log k n ) ) \mathrm{QP}_k = \mathrm{DTIME}(2^{O(\log^k n)}) QP k = DTIME ( 2 O ( l o g k n ) ) ; and one fixed computable law has class 2EXP . All these classes are independent of the confidence, and P = QP 1 ⊊ _1 \subsetneq 1 ⊊ QP 2 ⊊ ⋯ ⊊ _2 \subsetneq\cdots\subsetneq 2 ⊊ ⋯ ⊊ EXP ⊊ \subsetneq ⊊ 2EXP .
Further results. Positive drift and the nondegenerate regular zero-drift regimes give P (Theorems 3.3 and 3.6 ); a polynomial forward tail can overcome negative drift (Theorem 4.4). Computable iid laws collectively give REC , and one noncomputable law gives RE (Theorems 5.10 and 5.11). Common budgets over parameter sets depend on allowed-set infima (Appendix § 14 ). Figure 2 and Table 1 collect the regimes.
!The landscape of observation classes, showing the hierarchy from v a r n o t h i n g , S i g m a ∗ \\{\\varnothing,\\Sigma^*\\} v a r n o t hin g , S i g m a ∗ through RDEF, sparse laws, P = QP1 _1 1 , QP2 _2 2 , EXP, 2EXP, REC, and RE, with labels for frozen law, negative drift and all positive moments, one zero-drift class and a negative-drift antichain, positive drift / regular nondegenerate zero drift, μ [ k ] \mu_{[k]} μ [ k ] , μ E \mu_E μ E ; parity η β \eta_\beta η β ; gated η β g a t e \eta_\beta^{\mathrm{gate}} η β gate , K = 2 2 J K = 2^{2^J} K = 2 2 J ; squaring gates, computable laws (union), busy-beaver law, and forward-only gated clocks stay here bbox=0.19,0.29,0.82,0.64
Figure 2. The landscape of observation classes. Each sparse class lies strictly between RDEF and P ; the box asserts no ordering between the zero-drift class and the negative-drift antichain.
In Table 1, δ a \delta_a δ a denotes the point-mass probability law concentrated at a a a , including a = 0 a = 0 a = 0 and a = 1 a = 1 a = 1 . For an increment K K K , its positive and negative parts are K + = max ( K , 0 ) K^+ = \max(K, 0) K + = max ( K , 0 ) and K − = max ( − K , 0 ) K^- = \max(-K, 0) K − = max ( − K , 0 ) (Theorem 2.5). The deletion kernel D η D_\eta D η replaces backward moves by holds at their starting indices and keeps forward moves unchanged (Theorem 6.1).
Techniques and related work
Holding an inadmissible proposal at its starting position is a proposal-rejection rule, a standard way to construct Markov kernels [7 Section 3.2] . It differs from Lindley’s reflected update, which resets an overshoot below zero to zero [8 ] . The held walk dominates the reflected walk but is not monotone in its starting position (Theorem 2.9). Reflected-walk upper bounds therefore require a separate argument for the held rule.
The elementary drift regimes use the strong law, the central limit theorem, and moment bounds. Foster–Lyapunov methods [9 Chapters 11, 14 and 15] and exponential drift analysis [10 ] provide the standard framework for controlling negative-drift walks. Pemantle–Rosenthal [11 ] prove uniform moment bounds under negative drift, including a one-sided truncation version in their Corollary 3. Our proof of Theorem 4.1 gives the held-rule crossing estimate directly by truncating negative proposals in the conditional drift. It allows infinite negative mean. The subsequent identification with bounded-prefix recognition uses sequential locality, related to classical one-tape arguments [12 ] , rather than a further stability theorem.
Table 1. Selected fixed-confidence regimes, supplementing Figure 2. Each row holds for 0 < α < 1 0 < \alpha< 1 0 < α < 1 ; common-budget results are in Appendix F.
The relation between sums, maxima, and slowly varying tails is classical [13 , 14 ] . The independent-sum limit theorem in Croydon–Fribergh–Kumagai [15, Section 5] is a relevant probability comparison. Here elementary proposal-maxima bounds suffice for the explicit tails used in § 5 . The contribution is their realisation as fixed computable clocks with specified observation classes. The deterministic time hierarchy theorem supplies the strict separations between the familiar time classes [16 Section 3.1] .
The sparse constructions address a different realisation problem. Their explicit laws have crossing radii that alternate between effective fast and slow stages. Finite rational searches protect all previously chosen slow stages from future atoms in the zero-drift construction. A partition of tower scales gives the negative-drift antichain. Distinguishing input cells supply accepting-time lower bounds, while specific length languages give witnesses above the bounded-prefix class. Thus the class comparisons are proved for the constructed laws; an arbitrary resource-function hierarchy alone does not provide those probabilistic realisations.
Restart strategies for Las Vegas algorithms use independent new runs [17 ] . Our phase-reuse argument instead changes accessibility within the clock on one fixed deterministic path. The deletion coupling compares forward-displacement laws, and the gates compare all forward-only members of specified kernel families. In the sparse gated examples, density and counting orders concern the support of backward indices in index space. They do not measure the frequency of backward moves along a trajectory. Vanishing probabilities are specified at the doubling landing indices (Theorem 6.16).
The fuzzy-time computation background is described in § 1.1 .
Organisation
Section 2 defines the model and reachability dictionary; Sections 3 to 5 classify iid clocks, and Section 6 treats backward moves. Section 7 lists open problems. The appendices contain the sparse constructions, effectiveness proofs, calibrations, and common-budget results.
Model and the reachability dictionary
The jumpy observation model describes an observer who sees the configuration at a randomly chosen index, with consecutive indices chosen by a random clock that may jump. We define the computation path, clock, and acceptance rule. The access convention and sequential locality connect crossing estimates to language classes.
Machines, languages, and classes
The machine reads its input sequentially, which constrains clocks with small reach. We fix the machine conventions and recall the complexity classes used below.
A language over a finite alphabet Σ \Sigma Σ is a set L L L of finite input words; a machine decides L L L when it halts and accepts exactly the words of L L L . Polynomial time, written P , is the class of languages decidable with the number of executed transitions bounded by a polynomial in the input length. Exponential time is EXP = ⋃ k ≥ 1 DTIME ( 2 n k ) = \bigcup_{k \ge1} \operatorname{DTIME}(2^{n^k}) = ⋃ k ≥ 1 DTIME ( 2 n k ) , with the usual asymptotic running-time convention. Equivalently, a language has a deterministic decider with bound 2 p ( n ) 2^{p(n)} 2 p ( n ) for one fixed polynomial p p p , after enlarging the bound to cover every input length. Doubly exponential time is 2EXP = ⋃ k ≥ 1 DTIME ( 2 2 n k ) = \bigcup_{k \ge1} \operatorname{DTIME}(2^{2^{n^k}}) = ⋃ k ≥ 1 DTIME ( 2 2 n k ) . As for EXP , a language belongs to 2EXP exactly when it has a deterministic decider with bound 2 2 p ( n ) 2^{2^{p(n)}} 2 2 p ( n ) for one fixed polynomial p p p . For an integer k ≥ 1 k \ge1 k ≥ 1 put Q P k = ⋃ c ≥ 1 DTIME ( 2 c ( log 2 ( n + 2 ) ) k ) \mathrm{QP}_k = \bigcup_{c \ge1} \operatorname{DTIME}(2^{c(\log_2(n+2))^k}) QP k = ⋃ c ≥ 1 DTIME ( 2 c ( l o g 2 ( n + 2 ) ) k ) . Thus Q P 1 = P \mathrm{QP}_1 = \mathrm{P} QP 1 = P , and every Q P k \mathrm{QP}_k QP k is contained in EXP . The decidable languages are denoted REC , and the semidecidable (recursively enumerable) languages, accepted by machines that halt on yes-instances but need not halt on no-instances, are denoted RE . Thus P ⊆ R E C ⊊ R E \mathrm{P} \subseteq\mathrm{REC} \subsetneq\mathrm{RE} P ⊆ REC ⊊ RE . For example, over { 0 , 1 } \{0,1\} { 0 , 1 } , the prefix language 101 Σ ∗ 101\Sigma^* 101 Σ ∗ is decided in three transitions, while L e v e n = { x : ∣ x ∣ is even } L_{\mathrm{even}} = \{x : |x| \text{ is even}\} L even = { x : ∣ x ∣ is even } requires a full input scan. The reverse-definite languages RDEF , defined below, depend on only a bounded initial segment.
Definition 2.1 (Machine and logical path) Fix a finite nonempty input alphabet Σ \Sigma Σ . A machine has a finite tape alphabet Γ ⊇ Σ \Gamma\supseteq\Sigma Γ ⊇ Σ and a blank b ∈ Γ ∖ Σ b \in\Gamma\setminus\Sigma b ∈ Γ ∖ Σ . Its tape has cells 1 , 2 , … 1,2,\ldots 1 , 2 , … . Initially it contains x ∈ Σ ∗ x \in\Sigma^* x ∈ Σ ∗ in cells 1 , … , ∣ x ∣ 1,\ldots,|x| 1 , … , ∣ x ∣ and blanks elsewhere. The head starts at cell 1 in a fixed state q 0 q_0 q 0 , independent of x x x . A transition reads and writes the scanned cell, changes state, and moves the head by − 1 -1 − 1 , 0 0 0 , or + 1 +1 + 1 ; a left move at cell 1 leaves the head there. Input access is sequential. There is no input-length oracle or free access to an end marker.
The transition function is deterministic and defined on every nonhalting state and tape symbol. There are distinct accepting and rejecting halting states q a c c , q r e j q_{\mathrm{acc}},q_{\mathrm{rej}} q acc , q rej . We allow q 0 q_0 q 0 to be either halting state; this means that the machine halts before reading any symbol. A configuration records the state, head position, and full tape content. For each input x x x , let
ρ M ( x ) = ( C 0 x , C 1 x , … ) \rho_M(x) = (C_0^x,C_1^x,\ldots) ρ M ( x ) = ( C 0 x , C 1 x , … )
be the logical path obtained by executing transitions and repeating a halting configuration forever. The first halting index is T M ( x ) = inf { i ≥ 0 : C i x is halting } T_M(x) = \inf\{i \ge0 : C_i^x \text{ is halting}\} T M ( x ) = inf { i ≥ 0 : C i x is halting } , with inf ∅ = ∞ \inf\varnothing= \infty inf ∅ = ∞ . The first accepting index is A M ( x ) = T M ( x ) A_M(x) = T_M(x) A M ( x ) = T M ( x ) if this halt accepts, and A M ( x ) = ∞ A_M(x) = \infty A M ( x ) = ∞ otherwise. Thus a rejecting halt has finite T M ( x ) T_M(x) T M ( x ) but infinite A M ( x ) A_M(x) A M ( x ) ; a nonhalting path has both indices infinite.
We use N = { 0 , 1 , 2 , … } \mathbb{N} = \{0,1,2,\ldots\} N = { 0 , 1 , 2 , … } . Budgets are polynomials Q : N → N Q : \mathbb{N} \to\mathbb{N} Q : N → N with nonnegative integer coefficients, including the zero polynomial. Every polynomial upper bound has a majorant of this form on N \mathbb{N} N . We make the same enlargement for polynomial exponents in exponential-time bounds, taking p ( 0 ) ≥ 1 p(0) \ge1 p ( 0 ) ≥ 1 when needed. Machine-model scope is discussed in Theorem 2.19.
The held walk and the access rule
The held walk defines an independent-proposal random clock , or fuzzy clock . One tick supplies a configuration under the following access rule.
Definition 2.2 (Law and held walk) Fix one probability law μ \mu μ on Z \mathbb{Z} Z , independently of the input and the machine. Let K 1 , K 2 , … K_1,K_2,\ldots K 1 , K 2 , … be independent with law μ \mu μ . The logical position at tick t t t is X t X_t X t , where
X 0 = 0 , X t + 1 = { X t + K t + 1 , X t + K t + 1 ≥ 0 , X t , X t + K t + 1 < 0. (1) X_0=0,\qquad X_{t+1}=
\begin{cases}
X_t+K_{t+1}, & X_t+K_{t+1}\ge0,\\
X_t, & X_t+K_{t+1}<0.
\end{cases} \tag*{(1)} X 0 = 0 , X t + 1 = { X t + K t + 1 , X t , X t + K t + 1 ≥ 0 , X t + K t + 1 < 0. (1) An inadmissible negative jump is held at its starting position. This differs from the reflected update max ( 0 , X t + K t + 1 ) \max(0,X_t+K_{t+1}) max ( 0 , X t + K t + 1 ) ; the held rule is not generally order-preserving between starting positions. For an integer target T ≥ 0 T\ge0 T ≥ 0 , put
τ T = inf { t ≥ 0 : X t ≥ T } , inf ∅ = ∞ . \tau_T=\inf\{t\ge0:X_t\ge T\},\qquad\inf\varnothing=\infty. τ T = inf { t ≥ 0 : X t ≥ T } , inf ∅ = ∞. In the random-walk literature, τ T \tau_T τ T is the first passage time into [ T , ∞ ) [T,\infty) [ T , ∞ ) . We write Pr i \operatorname{Pr}_i Pr i for the same held walk started at i ≥ 0 i\ge0 i ≥ 0 , and omit the subscript when i = 0 i=0 i = 0 .
Conventions. Configurations are supplied at unit access cost, and the observer tests whether the state accepts. Configuration generation, proposal sampling, and index writing are excluded from that cost. Laws and parameters are fixed in advance, with each parameter fixed throughout a run. Witnesses and budget constants may depend on the language, clock, and confidence. No effectiveness is assumed unless stated. These observation resources make no claim of a physical implementation or ordinary machine speedup.
Classical complexity charges transitions; the present access convention charges observations.
Definition 2.3 (One-sided acceptance) A budget of Q Q Q ticks observes the configurations at ticks 0 , … , Q 0,\ldots,Q 0 , … , Q : there are Q + 1 Q+1 Q + 1 observations, including the free initial configuration. Their indices can repeat. For 0 < α < 1 0<\alpha<1 0 < α < 1 , a language L ⊆ Σ ∗ L\subseteq\Sigma^* L ⊆ Σ ∗ belongs to P μ , α ∗ \mathrm{P}^{*}_{\mu,\alpha} P μ , α ∗ if there are a deterministic machine M M M and a polynomial Q Q Q such that
for every x ∉ L x\notin L x ∈ / L , A M ( x ) = ∞ A_M(x)=\infty A M ( x ) = ∞ ;
for every x ∈ L x\in L x ∈ L , A M ( x ) < ∞ A_M(x)<\infty A M ( x ) < ∞ and Pr ( τ A M ( x ) ≤ Q ( ∣ x ∣ ) ) ≥ α \operatorname{Pr}(\tau_{A_M(x)}\le Q(|x|))\ge\alpha Pr ( τ A M ( x ) ≤ Q ( ∣ x ∣ )) ≥ α .
Set P μ ∗ = ⋂ 0 < α < 1 P μ , α ∗ \mathrm{P}^{*}_{\mu}=\bigcap_{0<\alpha<1}\mathrm{P}^{*}_{\mu,\alpha} P μ ∗ = ⋂ 0 < α < 1 P μ , α ∗ . Its quantifiers are ∀ α ∃ M ∃ Q ∀ x \forall\alpha\,\exists M\,\exists Q\,\forall x ∀ α ∃ M ∃ Q ∀ x : each confidence may receive its own witness, and no single witness is required to serve all levels.
Equalities at each fixed confidence pass to the intersection class.
Kernel classes. For a time-homogeneous stochastic kernel η \eta η on N \mathbb{N} N , fixed independently of the machine and input and started at zero, define C η , α \mathrm{C}_{\eta,\alpha} C η , α by Theorem 2.3 with the kernel’s hitting probability in place of the held law’s probability. Set C η = ⋂ 0 < α < 1 C η , α \mathrm{C}_{\eta}=\bigcap_{0<\alpha<1}\mathrm{C}_{\eta,\alpha} C η = ⋂ 0 < α < 1 C η , α . One tick supplies the configuration at the current index under the same access convention. For η = η μ \eta=\eta_\mu η = η μ in (3), these classes are P μ , α ∗ \mathrm{P}^{*}_{\mu,\alpha} P μ , α ∗ and P μ ∗ \mathrm{P}^{*}_{\mu} P μ ∗ .
Write ℓ α = − ln ( 1 − α ) \ell_\alpha=-\ln(1-\alpha) ℓ α = − ln ( 1 − α ) for the confidence logarithm; ln \ln ln is the natural logarithm.
On an accepting input, the accepting indices form the tail [ A M ( x ) , ∞ ) ∩ N [A_M(x),\infty)\cap\mathbb{N} [ A M ( x ) , ∞ ) ∩ N . Acceptance by tick q q q records a visit to this tail at some tick 0 ≤ t ≤ q 0\le t\le q 0 ≤ t ≤ q , even if the clock later jumps below A M ( x ) A_M(x) A M ( x ) . Soundness applies to the entire path of each no-instance, including nonhalting paths.
Lemma 2.4 (Reachable logical time) Fix a held law or a stochastic kernel on N \mathbb{N} N , started at zero, and 0 < α < 1 0<\alpha<1 0 < α < 1 . For every q ∈ N q\in\mathbb{N} q ∈ N , the quantity
R α ( q ) = max { T ∈ N : Pr ( τ T ≤ q ) ≥ α } R_\alpha(q)=\max\{T\in\mathbb{N}:\operatorname{Pr}(\tau_T\le q)\ge\alpha\} R α ( q ) = max { T ∈ N : Pr ( τ T ≤ q ) ≥ α } is well defined and finite. It is nondecreasing in q q q , and R α ( 0 ) = 0 R_\alpha(0)=0 R α ( 0 ) = 0 . A language belongs to the fixed-confidence class exactly when some deterministic recognizer M M M accepts precisely that language and some allowed polynomial Q Q Q satisfies
A M ( x ) ≤ R α ( Q ( ∣ x ∣ ) ) ( x ∈ L ) . A_M(x)\le R_\alpha(Q(|x|))\qquad(x\in L). A M ( x ) ≤ R α ( Q ( ∣ x ∣ )) ( x ∈ L ) . Here A M ( x ) A_M(x) A M ( x ) is finite on L L L and infinite off $L.
Proof. For finite q q q , the running maximum Z q = max 0 ≤ t ≤ q X t Z_q=\max_{0\le t\le q}X_t Z q = max 0 ≤ t ≤ q X t is surely finite: each of the finitely many positions takes its value in N \mathbb{N} N . Thus Pr ( τ T ≤ q ) = Pr ( Z q ≥ T ) \Pr(\tau_T\le q)=\Pr(Z_q\ge T) Pr ( τ T ≤ q ) = Pr ( Z q ≥ T ) decreases to zero as the integer T T T increases. Target zero is always eligible. The eligible targets therefore form a nonempty finite initial segment of N \mathbb{N} N , whose maximum is R α ( q ) R_\alpha(q) R α ( q ) . Increasing q q q can only enlarge this segment. At tick zero the only eligible target is zero.
For a recognizer accepting exactly L L L , one-sided soundness already holds off L L L . On each yes-instance, the initial-segment property gives Pr ( τ A M ( x ) ≤ Q ( ∣ x ∣ ) ) ≥ α \Pr(\tau_{A_M(x)}\le Q(|x|))\ge\alpha Pr ( τ A M ( x ) ≤ Q ( ∣ x ∣ )) ≥ α if and only if A M ( x ) ≤ R α ( Q ( ∣ x ∣ ) ) A_M(x)\le R_\alpha(Q(|x|)) A M ( x ) ≤ R α ( Q ( ∣ x ∣ )) .
For a fixed clock and confidence, define the crossing budget by
q α ( T ) = min { q ∈ N 0 : Pr ( τ T ≤ q ) ≥ α } , min ∅ = ∞ . (2) q_\alpha(T)=\min\{q\in\mathbb{N}_0:\Pr(\tau_T\le q)\ge\alpha\},\qquad\min\varnothing=\infty. \tag*{(2)} q α ( T ) = min { q ∈ N 0 : Pr ( τ T ≤ q ) ≥ α } , min ∅ = ∞. (2)
The quantities used throughout the paper have the following meanings.
Table 1 (PDF p. 8).
The budget and radius are inverse views of the same crossing probabilities: q α ( T ) ≤ q q_\alpha(T)\le q q α ( T ) ≤ q exactly when T ≤ R α ( q ) T\le R_\alpha(q) T ≤ R α ( q ) . Theorem 2.18 translates radius growth into class bounds.
Definition 2.5 (Moments and drift) For K ∼ μ K\sim\mu K ∼ μ , put K + = max ( K , 0 ) K^+=\max(K,0) K + = max ( K , 0 ) and K − = max ( − K , 0 ) K^-=\max(-K,0) K − = max ( − K , 0 ) . For r > 0 r>0 r > 0 , the positive moment m r = E [ ( K + ) r ] m_r=\mathbb{E}[(K^+)^r] m r = E [( K + ) r ] is allowed to be infinite. We use the finite drift v = E K v=\mathbb{E}K v = E K only when E ∣ K ∣ = E K + + E K − < ∞ \mathbb{E}|K|=\mathbb{E}K^++\mathbb{E}K^-<\infty E ∣ K ∣ = E K + + E K − < ∞ . Under this condition the variance is E [ ( K − v ) 2 ] \mathbb{E}[(K-v)^2] E [( K − v ) 2 ] , possibly infinite. Otherwise an extended mean may be + ∞ +\infty + ∞ or − ∞ -\infty − ∞ ; if both parts are infinite, E K \mathbb{E}K E K is undefined. None of these cases is assigned a finite drift. A positive exponential moment means E e θ K < ∞ \mathbb{E}e^{\theta K}<\infty E e θ K < ∞ for some θ > 0 \theta>0 θ > 0 . Since K K K is integer valued, Pr ( K > 0 ) > 0 \Pr(K>0)>0 Pr ( K > 0 ) > 0 holds exactly when μ ( k 0 ) > 0 \mu(k_0)>0 μ ( k 0 ) > 0 for some integer k 0 ≥ 1 k_0\ge1 k 0 ≥ 1 .
Definition 2.6 (Computable law) A probability law μ \mu μ on Z \mathbb{Z} Z is computable if an algorithm, on input ( k , p ) ∈ Z × N (k,p)\in\mathbb{Z}\times\mathbb{N} ( k , p ) ∈ Z × N , returns a rational d ( k , p ) d(k,p) d ( k , p ) with ∣ d ( k , p ) − μ ( k ) ∣ ≤ 2 − p |d(k,p)-\mu(k)|\le2^{-p} ∣ d ( k , p ) − μ ( k ) ∣ ≤ 2 − p . The law has total mass one, as usual. Thus the point probabilities are computable uniformly in k k k , with certified errors. No bound on the running time of this algorithm is part of the definition.
This agrees with the standard notion of a computable probability measure on the discrete integers; Appendix C explains the equivalence and derives an effective tail bound and a decidability result.
The transition kernel and two basic bounds
The kernel identifies the held boundary rule, and two bounds compare its reach with free proposals. A Markov kernel on N \mathbb{N} N is a matrix of nonnegative transition probabilities whose rows sum to one. For (1) it is
η μ ( i , j ) = { μ ( j − i ) , j ≠ i , μ ( 0 ) + Pr ( K < − i ) , j = i . (3) \eta_\mu(i,j)=
\begin{cases}
\mu(j-i), & j\ne i,\\
\mu(0)+\Pr(K<-i), & j=i.
\end{cases} \tag*{(3)} η μ ( i , j ) = { μ ( j − i ) , μ ( 0 ) + Pr ( K < − i ) , j = i , j = i . (3)
The row sum is ∑ k ≥ − i μ ( k ) + Pr ( K < − i ) = 1 \sum_{k\ge-i}\mu(k)+\Pr(K<-i)=1 ∑ k ≥ − i μ ( k ) + Pr ( K < − i ) = 1 . Off the diagonal, the transition probability depends only on the displacement; conversely, this property characterizes the held kernels among stochastic kernels, with a unique recovering law (Theorem 2.7).
Proof. Necessity follows from (3). For sufficiency, row sums give ∑ k ≥ − i , k ≠ 0 φ ( k ) ≤ 1 \sum_{k\ge-i,k\ne0}\varphi(k) \le1 ∑ k ≥ − i , k = 0 φ ( k ) ≤ 1 ; monotone convergence gives ∑ k ≠ 0 φ ( k ) ≤ 1 \sum_{k\ne0}\varphi(k) \le1 ∑ k = 0 φ ( k ) ≤ 1 [18 Theorem 1.5.7] . Set μ ( k ) = φ ( k ) \mu(k) = \varphi(k) μ ( k ) = φ ( k ) for k ≠ 0 k \ne0 k = 0 and μ ( 0 ) = 1 − ∑ k ≠ 0 φ ( k ) \mu(0) = 1-\sum_{k\ne0}\varphi(k) μ ( 0 ) = 1 − ∑ k = 0 φ ( k ) ; stochasticity determines η ( i , i ) = μ ( 0 ) + ∑ k < − i μ ( k ) \eta(i,i) = \mu(0) + \sum_{k<-i}\mu(k) η ( i , i ) = μ ( 0 ) + ∑ k < − i μ ( k ) . For k ≠ 0 k \ne0 k = 0 , any row i ≥ max ( 0 , − k ) i \ge\max(0,-k) i ≥ max ( 0 , − k ) recovers μ ( k ) = η ( i , i + k ) \mu(k) = \eta(i,i+k) μ ( k ) = η ( i , i + k ) , and normalization recovers μ ( 0 ) \mu(0) μ ( 0 ) , proving uniqueness; a row i ≥ w i \ge w i ≥ w recovers every mass of a law supported in [ − w , w ] [-w,w] [ − w , w ] .
The hold rule never moves farther down than an executed negative proposal. It also never gains more than the sum of the positive proposals.
Lemma 2.8 (Comparison with the free walk) With S t = ∑ j = 1 t K j S_t = \sum_{j=1}^{t}K_j S t = ∑ j = 1 t K j , a held walk started at i ≥ 0 i \ge0 i ≥ 0 satisfies X t ≥ i + S t X_t \ge i+S_t X t ≥ i + S t for every t t t . Started at zero, it also satisfies
max 0 ≤ s ≤ q X s ≤ ∑ j = 1 q K j + . (4) \max_{0\le s\le q} X_s \le\sum_{j=1}^{q}K_j^{+}. \tag*{(4)} 0 ≤ s ≤ q max X s ≤ j = 1 ∑ q K j + . (4) Consequently Pr i ( τ T ≤ q ) ≥ Pr ( S q ≥ T ) \operatorname{Pr}_i(\tau_T \le q) \ge\operatorname{Pr}(S_q \ge T) Pr i ( τ T ≤ q ) ≥ Pr ( S q ≥ T ) for i ≥ 0 i \ge0 i ≥ 0 .
Proof. For the lower bound we use induction on t t t ; the case t = 0 t=0 t = 0 is X 0 = i = i + S 0 X_0=i=i+S_0 X 0 = i = i + S 0 . If the proposal K t + 1 K_{t+1} K t + 1 is executed, then X t + 1 = X t + K t + 1 ≥ i + S t + K t + 1 = i + S t + 1 X_{t+1}=X_t+K_{t+1}\ge i+S_t+K_{t+1}=i+S_{t+1} X t + 1 = X t + K t + 1 ≥ i + S t + K t + 1 = i + S t + 1 . If it is held, then X t + K t + 1 < 0 ≤ X t X_t+K_{t+1}<0\le X_t X t + K t + 1 < 0 ≤ X t , so K t + 1 < 0 K_{t+1}<0 K t + 1 < 0 and X t + 1 = X t ≥ i + S t > i + S t + 1 X_{t+1}=X_t\ge i+S_t>i+S_{t+1} X t + 1 = X t ≥ i + S t > i + S t + 1 .
For the upper bound, both cases give X t + 1 ≤ X t + K t + 1 + X_{t+1}\le X_t+K_{t+1}^{+} X t + 1 ≤ X t + K t + 1 + . Started at zero, this yields X s ≤ ∑ j = 1 s K j + X_s\le\sum_{j=1}^{s}K_j^{+} X s ≤ ∑ j = 1 s K j + for every s s s . The right-hand side is nondecreasing in s s s , which proves (4).
Finally, with common proposals, S q ≥ T S_q\ge T S q ≥ T implies X q ≥ i + S q ≥ T X_q\ge i+S_q\ge T X q ≥ i + S q ≥ T , hence τ T ≤ q \tau_T\le q τ T ≤ q . This inclusion of events gives Pr i ( τ T ≤ q ) ≥ Pr ( S q ≥ T ) \operatorname{Pr}_i(\tau_T\le q)\ge\operatorname{Pr}(S_q\ge T) Pr i ( τ T ≤ q ) ≥ Pr ( S q ≥ T ) .
Lemma 2.10 (Elementary proposal-maxima comparison) For iid held proposals and q ≥ 1 q\ge1 q ≥ 1 , put J q = max 1 ≤ j ≤ q K j + J_q=\max_{1\le j\le q}K_j^{+} J q = max 1 ≤ j ≤ q K j + . For integer T ≥ 1 T\ge1 T ≥ 1 ,
{ J q ≥ T } ⊆ { τ T ≤ q } ⊆ { J q ≥ T / q } . \{J_q\ge T\}\subseteq\{\tau_T\le q\}\subseteq\{J_q\ge T/q\}. { J q ≥ T } ⊆ { τ T ≤ q } ⊆ { J q ≥ T / q } . The integer quantile u α ( q ) = max { u ∈ N : Pr ( J q ≥ u ) ≥ α } u_\alpha(q)=\max\{u\in\mathbb{N}:\operatorname{Pr}(J_q\ge u)\ge\alpha\} u α ( q ) = max { u ∈ N : Pr ( J q ≥ u ) ≥ α } is finite and satisfies u α ( q ) ≤ R α ( q ) ≤ q u α ( q ) u_\alpha(q)\le R_\alpha(q)\le qu_\alpha(q) u α ( q ) ≤ R α ( q ) ≤ q u α ( q ) . For q = 0 q=0 q = 0 set u α ( 0 ) = 0 u_\alpha(0)=0 u α ( 0 ) = 0 , agreeing with R α ( 0 ) = 0 R_\alpha(0)=0 R α ( 0 ) = 0 . For q ≥ 1 q\ge1 q ≥ 1 and every real z > 0 z>0 z > 0 , define F ‾ ( z ) = Pr ( K + ≥ z ) \overline{F}(z)=\operatorname{Pr}(K^{+}\ge z) F ( z ) = Pr ( K + ≥ z ) . Then
Pr ( J q ≥ z ) = 1 − ( 1 − F ‾ ( z ) ) q ; Pr ( J q ≥ z ) ≥ α ⟺ F ‾ ( z ) ≥ 1 − ( 1 − α ) 1 / q . \operatorname{Pr}(J_q\ge z)=1-(1-\overline{F}(z))^q;\qquad\operatorname{Pr}(J_q\ge z)\ge\alpha\ \Longleftrightarrow\ \overline{F}(z)\ge1-(1-\alpha)^{1/q}. Pr ( J q ≥ z ) = 1 − ( 1 − F ( z ) ) q ; Pr ( J q ≥ z ) ≥ α ⟺ F ( z ) ≥ 1 − ( 1 − α ) 1/ q . The tail threshold is real; the quantile remains over integer targets.
Proof. A proposal at least T T T crosses from every nonnegative position. Conversely, (4) gives max s ≤ q X s ≤ ∑ j ≤ q K j + ≤ q J q \max_{s\le q}X_s\le\sum_{j\le q}K_j^{+}\le qJ_q max s ≤ q X s ≤ ∑ j ≤ q K j + ≤ q J q . The maximum J q J_q J q is finite almost surely, so its eligible integer targets form a nonempty finite initial segment, as in Theorem 2.4. If T T T is reachable at confidence α \alpha α , then ⌈ T / q ⌉ \lceil T/q\rceil ⌈ T / q ⌉ is eligible for J q J_q J q , hence T ≤ q u α ( q ) T\le qu_\alpha(q) T ≤ q u α ( q ) . The lower quantile bound follows from the first inclusion, also when u α ( q ) = 0 u_\alpha(q)=0 u α ( q ) = 0 . Independence gives the tail formula and threshold.
The factor q q q can matter: for δ 1 \delta_1 δ 1 , u α ( q ) = 1 u_\alpha(q)=1 u α ( q ) = 1 but R α ( q ) = q R_\alpha(q)=q R α ( q ) = q . This estimate alone does not classify arbitrary clocks.
Lemma 2.11 (Supported forward blocks) If p 0 = μ ( k 0 ) > 0 p_0=\mu(k_0)>0 p 0 = μ ( k 0 ) > 0 for some k 0 ≥ 1 k_0\ge1 k 0 ≥ 1 , then for T ≥ 1 T\ge1 T ≥ 1 , m = ⌈ T / k 0 ⌉ m=\lceil T/k_0\rceil m = ⌈ T / k 0 ⌉ , j ≥ 1 j\ge1 j ≥ 1 , and every i ≥ 0 i\ge0 i ≥ 0 ,
Pr i ( τ T > j m ) ≤ ( 1 − p 0 m ) j ≤ e − j p 0 m . \Pr_i(\tau_T>jm)\le(1-p_0^m)^j\le e^{-jp_0^m}. i Pr ( τ T > j m ) ≤ ( 1 − p 0 m ) j ≤ e − j p 0 m . Thus every fixed target has a constant tick budget at any fixed confidence.
Proof. Partition the proposals into j j j disjoint blocks of length m m m . The event that all proposals in a block equal k 0 k_0 k 0 has probability p 0 m p_0^{m} p 0 m . The independent-trial formula gives failure probability ( 1 − p 0 m ) j (1-p_0^m)^j ( 1 − p 0 m ) j . Each such block reaches T T T from any nonnegative starting position, since all its jumps execute and m k 0 ≥ T mk_0\ge T m k 0 ≥ T . If no hit occurs, all these events fail. For confidence α \alpha α , take j = ⌈ p 0 − m ℓ α ⌉ j=\lceil p_0^{-m}\ell_\alpha\rceil j = ⌈ p 0 − m ℓ α ⌉ .
For nonnegative proposals, equality holds in (4) . For a law on { 0 , 1 } \{0,1\} { 0 , 1 } with μ ( 1 ) = p > 0 \mu(1)=p>0 μ ( 1 ) = p > 0 and target 1 1 1 , the block failure probability is ( 1 − p ) j (1-p)^j ( 1 − p ) j from zero. Larger-target block budgets can exceed the optimal budget.
Proposition 2.12 (Frozen laws) If Pr ( K > 0 ) = 0 \Pr(K>0)=0 Pr ( K > 0 ) = 0 , then X t = 0 X_t=0 X t = 0 almost surely for every t t t and P μ , α = P μ ∗ = { ∅ , Σ ∗ } \mathcal{P}_{\mu,\alpha}=\mathcal{P}^*_\mu=\{\varnothing,\Sigma^*\} P μ , α = P μ ∗ = { ∅ , Σ ∗ } for every 0 < α < 1 0<\alpha<1 0 < α < 1 .
Proof. Every proposal at zero is either zero or held. A yes-instance must therefore have A M ( x ) = 0 A_M(x)=0 A M ( x ) = 0 . The initial state is independent of x x x : if it accepts, soundness forces L = Σ ∗ L=\Sigma^* L = Σ ∗ ; otherwise completeness forces L = ∅ L=\varnothing L = ∅ . The immediately accepting and immediately rejecting machines attain both languages.
Sequential locality and the reachability dictionary
This is the standard sequential-input locality observation: a head moving by at most one cell per transition cannot read beyond the first T T T cells in T T T transitions; related crossing-sequence arguments for one-tape machines go back to [12 ] . We give the proof for our tape convention.
Lemma 2.13 (Prefix truncation) For a deterministic machine and an integer T ≥ 0 T\ge0 T ≥ 0 , let u u u be the prefix of x x x of length min ( ∣ x ∣ , T ) \min(|x|,T) min ( ∣ x ∣ , T ) . The runs on x x x and u u u have the same states and head positions through index T T T . If either run first halts at an index at most T T T , the other first halts at the same index and in the same state.
Proof. If ∣ x ∣ ≤ T |x|\le T ∣ x ∣ ≤ T , then u = x u=x u = x . Otherwise the initial tapes agree in cells 1 , … , T 1,\ldots,T 1 , … , T . Before transition t + 1 t+1 t + 1 , for 0 ≤ t < T 0\le t<T 0 ≤ t < T , the head is at most at cell t + 1 ≤ T t+1\le T t + 1 ≤ T . Induction shows that the state, head position, and contents of these cells agree through index T T T . Cell T + 1 T+1 T + 1 may be the final head position but is not read during these T T T transitions. In the truncated input its blank therefore causes no earlier difference. The case T = 0 T=0 T = 0 compares initial states without reading a cell. Equality of states through T T T gives the assertion about first halting.
Lemma 2.14 (Sublinear accepting time is constant) Let g : N → [ 0 , ∞ ) g:\mathbb{N}\to[0,\infty) g : N → [ 0 , ∞ ) be nondecreasing and satisfy g ( n ) / n → 0 g(n)/n\to0 g ( n ) / n → 0 as n → ∞ n\to\infty n → ∞ . Under the sequential tape conventions of Theorem 2.1, suppose M M M accepts exactly L L L and A M ( x ) ≤ g ( ∣ x ∣ ) A_M(x)\le g(|x|) A M ( x ) ≤ g ( ∣ x ∣ ) for every x ∈ L x\in L x ∈ L . Then A M ( x ) ≤ t l o c A_M(x)\le t_{\mathrm{loc}} A M ( x ) ≤ t loc on L L L for one integer t l o c t_{\mathrm{loc}} t loc depending only on g g g . If M M M is total and T M ( x ) ≤ g ( ∣ x ∣ ) T_M(x)\le g(|x|) T M ( x ) ≤ g ( ∣ x ∣ ) on every input, its total halting time is likewise bounded by t l o c t_{\mathrm{loc}} t loc .
Proof. For x ∈ L x \in L x ∈ L put T = A M ( x ) T = A_M(x) T = A M ( x ) and truncate it to u u u of length min ( ∣ x ∣ , T ) \min(|x|, T) min ( ∣ x ∣ , T ) . By Theorem 2.13, M M M first accepts u u u at index T T T . Soundness implies u ∈ L u \in L u ∈ L , so
T = A M ( u ) ≤ g ( ∣ u ∣ ) ≤ g ( T ) . T = A_M(u) \le g(|u|) \le g(T). T = A M ( u ) ≤ g ( ∣ u ∣ ) ≤ g ( T ) .
Choose an integer R l o c ≥ 1 R_{\mathrm{loc}} \ge1 R loc ≥ 1 such that g ( t ) < t g(t) < t g ( t ) < t for every integer t ≥ R l o c t \ge R_{\mathrm{loc}} t ≥ R loc . Such an R l o c R_{\mathrm{loc}} R loc exists because g ( t ) / t → 0 g(t)/t \to0 g ( t ) / t → 0 . The displayed inequality forces T < R l o c T < R_{\mathrm{loc}} T < R loc , so t l o c = R l o c − 1 t_{\mathrm{loc}} = R_{\mathrm{loc}} - 1 t loc = R loc − 1 works for every accepting input. For a total machine use T = T M ( x ) T = T_M(x) T = T M ( x ) ; its bound applies to u u u regardless of the halting state.
The proof only needs g ( t ) < t g(t) < t g ( t ) < t eventually. If A M ( x ) ≤ c ∣ x ∣ + b A_M(x) \le c|x| + b A M ( x ) ≤ c ∣ x ∣ + b on accepting inputs with 0 ≤ c < 1 0 \le c < 1 0 ≤ c < 1 and b ≥ 0 b \ge0 b ≥ 0 , prefix truncation gives T ≤ c T + b T \le cT + b T ≤ c T + b , hence A M ( x ) ≤ ⌊ b / ( 1 − c ) ⌋ A_M(x) \le\left\lfloor b/(1-c) \right\rfloor A M ( x ) ≤ ⌊ b / ( 1 − c ) ⌋ . The same holds for total halting time. At the linear boundary a scan to the first blank takes ∣ x ∣ + 1 |x| + 1 ∣ x ∣ + 1 transitions, so unbounded times are possible.
Corollary 2.15 (Logarithmic accepting time) If M M M accepts exactly L L L and A M ( x ) ≤ c ( log 2 ( ∣ x ∣ + 1 ) + 1 ) A_M(x) \le c(\log_2(|x| + 1) + 1) A M ( x ) ≤ c ( log 2 ( ∣ x ∣ + 1 ) + 1 ) on L L L for a fixed c > 0 c > 0 c > 0 , then its accepting time is bounded by a constant depending only on c c c . The corresponding statement holds for total halting time bounded by this expression on every input.
Proof. Apply Theorem 2.14 to g ( n ) = c ( log 2 ( n + 1 ) + 1 ) g(n) = c(\log_2(n + 1) + 1) g ( n ) = c ( log 2 ( n + 1 ) + 1 ) , which is nonnegative, nondecreasing and o ( n ) o(n) o ( n ) .
Definition 2.16 (Reverse-definite language) A language is reverse definite if, for some fixed integer k ≥ 1 k \ge1 k ≥ 1 , A ⊆ Σ k A \subseteq\Sigma^k A ⊆ Σ k , and F ⊆ Σ < k F \subseteq\Sigma^{<k} F ⊆ Σ < k ,
L = F ∪ ⋃ v ∈ A v Σ ∗ . (5) L = F \cup\bigcup_{v \in A} v\Sigma^*. \tag*{(5)} L = F ∪ v ∈ A ⋃ v Σ ∗ . (5)
Write RDEF for this class. Membership depends on the first k k k symbols, with shorter words handled separately. This is the classical reverse-definite class [19 Section 4.2] . The prefix language 101 Σ ∗ 101\Sigma^* 101 Σ ∗ of Section 2 is reverse definite with k = 3 k = 3 k = 3 and A = { 101 } A = \{101\} A = { 101 } ; the even-length language is not, since for any k k k the words a 2 k a^{2k} a 2 k and a 2 k + 1 a^{2k+1} a 2 k + 1 share their length-k k k prefix and have opposite membership.
The original theory of definite events and automata is due to Brzozowski and to Perles, Rabin, and Shamir [20 , 21 ] . Definite events depend on a bounded suffix; reversing words gives the bounded-prefix counterpart used here.
Lemma 2.17 (Bounded accepting time and bounded prefixes) A language has a sound recognizer with a constant accepting-time bound if and only if it is reverse definite. Every reverse-definite language has a total constant-time decider. If a clock reaches every finite target at confidence α \alpha α with some finite budget, then RDEF ⊆ C α \subseteq C_\alpha ⊆ C α , where C α C_\alpha C α denotes its class. In particular this holds for any iid held law with Pr ( K > 0 ) > 0 \Pr(K > 0) > 0 Pr ( K > 0 ) > 0 .
Proof. Suppose M M M accepts exactly L L L by a constant index t 0 t_0 t 0 on all accepted inputs. Put k = max ( 1 , t 0 ) k = \max(1,t_0) k = max ( 1 , t 0 ) . Membership is equivalent to acceptance by index k k k . By Theorem 2.13, words of length at least k k k have the same outcome as their length-k k k prefixes. Thus (5) holds with A = L ∩ Σ k A = L \cap\Sigma^k A = L ∩ Σ k and F = L ∩ Σ < k F = L \cap\Sigma^{<k} F = L ∩ Σ < k . Conversely, for fixed k , A , F k,A,F k , A , F , a finite-control decider reads at most the first k k k cells. If it encounters a blank after j < k j < k j < k symbols, it decides membership in F F F in j + 1 ≤ k j + 1 \le k j + 1 ≤ k transitions. If it reads k k k symbols, it decides whether their prefix is in A A A in k k k transitions. This also handles the empty input by reading its first blank. Its time bound is constant. A finite budget reaching that constant target at confidence α \alpha α observes every accepting tail; no-instance soundness is inherited from the decider. For an iid law, Theorem 2.11 supplies the finite-target hypothesis.
Lemma 2.18 (Reachability dictionary) Fix a clock on N \mathbb{N} N , started at zero, and a confidence 0 < α < 1 0 < \alpha< 1 0 < α < 1 . Write C α C_\alpha C α for its class and R α ( q ) R_\alpha(q) R α ( q ) for Theorem 2.4’s radius. Then the following hold.
(i) If R α ( q ) ≤ C ( q + 1 ) d R_\alpha(q) \le C(q+1)^d R α ( q ) ≤ C ( q + 1 ) d for finite C , d > 0 C,d>0 C , d > 0 , then C α ⊆ P C_\alpha\subseteq\mathbf{P} C α ⊆ P .
(ii) If R α ( q ) ≥ c q ε R_\alpha(q) \ge cq^\varepsilon R α ( q ) ≥ c q ε eventually, with c , ε > 0 c,\varepsilon>0 c , ε > 0 , then P ⊆ C α \mathbf{P} \subseteq C_\alpha P ⊆ C α .
(iii) If 1 + R α ( q ) = O ( ( q + 1 ) ε ) 1+R_\alpha(q)=O((q+1)^\varepsilon) 1 + R α ( q ) = O (( q + 1 ) ε ) for every ε > 0 \varepsilon>0 ε > 0 , and every finite target is reachable at confidence α \alpha α with a finite budget, then C α = R D E F C_\alpha=\mathbf{RDEF} C α = RDEF .
(iv) If eventually 2 c q ε ≤ R α ( q ) ≤ 2 C ( q + 1 ) d 2^{cq^\varepsilon}\le R_\alpha(q)\le2^{C(q+1)^d} 2 c q ε ≤ R α ( q ) ≤ 2 C ( q + 1 ) d for positive c , ε , C , d c,\varepsilon,C,d c , ε , C , d , then C α = E X P C_\alpha=\mathbf{EXP} C α = EXP .
Proof. For a witness ( M , Q ) (M,Q) ( M , Q ) , Theorem 2.4 bounds its accepting indices by R α ( Q ( n ) ) R_\alpha(Q(n)) R α ( Q ( n )) . In (i), choose an integer polynomial majorant for C ( Q ( n ) + 1 ) d C(Q(n)+1)^d C ( Q ( n ) + 1 ) d . Simulate M M M through that logical index, checking index zero, and reject on timeout. Completeness ensures acceptance on every yes-instance, while soundness excludes acceptance on a no-instance. Reading the input length and polynomial simulation overhead preserve polynomial time. This gives (i). The upper estimate in (iv) gives an exponential-time decider in the same way. The finite number of exceptional budgets is absorbed in the bound using the finite radii.
For (ii), let H ( n ) H(n) H ( n ) be a polynomial logical-time bound of a decider. Choose an integer polynomial Q Q Q of sufficiently large degree and coefficient that c Q ( n ) ε ≥ H ( n ) cQ(n)^\varepsilon\ge H(n) c Q ( n ) ε ≥ H ( n ) and the eventual radius estimate applies. Enlarge its constant term for the finitely many small input lengths. The radius tends to infinity, so every fixed target also has a finite adequate budget. Theorem 2.4 now gives the inclusion, with the same sound decider. For the lower inclusion in (iv), use H ( n ) = 2 p ( n ) H(n)=2^{p(n)} H ( n ) = 2 p ( n ) and choose Q Q Q with c Q ( n ) ε ≥ p ( n ) cQ(n)^\varepsilon\ge p(n) c Q ( n ) ε ≥ p ( n ) . Any classical machine-model simulation overhead can be absorbed in p p p . This proves (iv).
For (iii), write Q ( n ) ≤ a ( n + 1 ) d Q(n)\le a(n+1)^d Q ( n ) ≤ a ( n + 1 ) d with d ≥ 1 d\ge1 d ≥ 1 and a ≥ 1 a\ge1 a ≥ 1 . Choose ε < 1 / d \varepsilon<1/d ε < 1/ d . The subpolynomial estimate gives a nondecreasing upper bound g ( n ) = O ( ( n + 1 ) d ε ) = o ( n ) g(n)=O((n+1)^{d\varepsilon})=o(n) g ( n ) = O (( n + 1 ) d ε ) = o ( n ) for R α ( Q ( n ) ) R_\alpha(Q(n)) R α ( Q ( n )) , after enlarging a constant to cover small budgets. By Theorem 2.14, every accepting time is then bounded by a constant. Theorem 2.17 identifies the language as reverse definite and, using the finite-target hypothesis, proves the lower inclusion. □
Remark 2.19 (Machine-model robustness) Prefix truncation holds for finitely many initially blank work tapes and sequential input heads: in T T T transitions each head reads only the first T T T input cells. Thus the RDEF results and the n + 1 n+1 n + 1 distinguishing-cell lower bound in Appendix § 9 remain valid without a length oracle. The P , QP k _k k , EXP , 2EXP , REC , and RE classifications are invariant under polynomially related sequential models [16 ] ; overheads are absorbed by radius bounds or halting envelopes, and for QP k _k k only change the constant. The sparse law’s stage budgets T j ′ T_j' T j ′ dominate every fixed polynomial slowdown, preserving its separation from L e v e n L_{\mathrm{even}} L even , although its exact class may depend on the machine model.
Independent jumps up to P
This section bounds crossing probabilities by forward moments and proves linear or quadratic crossing budgets under suitable drift assumptions. The moment bound gives containment in P ; the budget estimates give equality.
Theorem 3.1 (Positive-moment containment) If E [ ( K + ) r ] < ∞ \mathbb{E}[(K^+)^r]<\infty E [( K + ) r ] < ∞ for some r > 0 r>0 r > 0 , then P μ , α ∗ ⊆ P \mathbf{P}_{\mu,\alpha}^*\subseteq\mathbf{P} P μ , α ∗ ⊆ P for every 0 < α < 1 0<\alpha<1 0 < α < 1 . More precisely, for 0 < r ≤ 1 0<r\le1 0 < r ≤ 1 , m r = E [ ( K + ) r ] < ∞ m_r=\mathbb{E}[(K^+)^r]<\infty m r = E [( K + ) r ] < ∞ , T ≥ 1 T\ge1 T ≥ 1 , and q ≥ 0 q\ge0 q ≥ 0 ,
Pr ( τ T ≤ q ) ≤ q m r T r . (6) \Pr(\tau_T\le q)\le\frac{qm_r}{T^r}. \tag*{(6)} Pr ( τ T ≤ q ) ≤ T r q m r . (6) For such 0 < r ≤ 1 0<r\le1 0 < r ≤ 1 , a witness ( M , Q ) (M,Q) ( M , Q ) therefore has A M ( x ) ≤ ( m r Q ( ∣ x ∣ ) / α ) 1 / r A_M(x)\le(m_rQ(|x|)/\alpha)^{1/r} A M ( x ) ≤ ( m r Q ( ∣ x ∣ ) / α ) 1/ r on every yes-instance with A M ( x ) > 0 A_M(x)>0 A M ( x ) > 0 .
Proof. If a moment of order r > 1 r>1 r > 1 is finite, the first moment is finite because y ≤ 1 + y r y\le1+y^r y ≤ 1 + y r for y ≥ 0 y\ge0 y ≥ 0 . We may therefore assume 0 < r ≤ 1 0<r\le1 0 < r ≤ 1 . By (4) , subadditivity of the r r r th power, and Markov’s inequality [18 Theorem 1.6.4] ,
Pr ( τ T ≤ q ) ≤ E [ ( ∑ j = 1 q K j + ) r ] T r ≤ ∑ j = 1 q E [ ( K j + ) r ] T r = q m r T r . \Pr(\tau_T\le q)\le\frac{\mathbb{E}\left[\left(\sum_{j=1}^{q}K_j^+\right)^r\right]}{T^r}\le\frac{\sum_{j=1}^{q}\mathbb{E}\left[(K_j^+)^r\right]}{T^r}=\frac{qm_r}{T^r}. Pr ( τ T ≤ q ) ≤ T r E [ ( ∑ j = 1 q K j + ) r ] ≤ T r ∑ j = 1 q E [ ( K j + ) r ] = T r q m r .
Completeness gives the claimed bound when A M ( x ) > 0 A_M(x)>0 A M ( x ) > 0 . If m r = 0 m_r=0 m r = 0 , then K + = 0 K^+=0 K + = 0 almost surely, and Theorem 2.12 applies. For m r > 0 m_r>0 m r > 0 , (6) bounds the radius by ( m r q / α ) 1 / r (m_rq/\alpha)^{1/r} ( m r q / α ) 1/ r . Apply Theorem 2.18(i). □
A law with all positive moments infinite can still have class P ; see Appendix § 12 .
The sharpness of the moment-bound coefficient is illustrated in Appendix § 13 .§ 1 . Recall the crossing budget q α ( T ) q_\alpha(T) q α ( T ) from (2) . All asymptotic comparisons keep the law and confidence fixed and let the integer target T T T tend to infinity; their constants may depend on both. For μ = δ 1 \mu=\delta_1 μ = δ 1 , τ T = T \tau_T=T τ T = T from zero, so the linear scale in the next lemma cannot be improved in general.
Positive drift without a variance assumption
Positive drift makes the free walk advance linearly with high probability. The comparison lemma transfers that estimate to the held walk.
Lemma 3.2 (Linear crossing budget) If E ∣ K ∣ < ∞ \mathbb{E}|K|<\infty E ∣ K ∣ < ∞ and v = E K > 0 v=\mathbb{E}K>0 v = E K > 0 , then for each 0 < α < 1 0<\alpha<1 0 < α < 1 there is an integer c α , d r i f t ≥ 1 c_{\alpha,\mathrm{drift}}\ge1 c α , drift ≥ 1 such that, for every T ≥ 1 T\ge1 T ≥ 1 and i ≥ 0 i\ge0 i ≥ 0 ,
Pr i ( τ T ≤ c α , d r i f t T ) ≥ α . \Pr_i(\tau_T\le c_{\alpha,\mathrm{drift}}T)\ge\alpha. i Pr ( τ T ≤ c α , drift T ) ≥ α .
Proof. The strong law of large numbers gives S q / q → v S_q/q\to v S q / q → v almost surely, and hence in probability [18 Theorem 2.4.1] . Its integrability and independence hypotheses hold for the increments. Choose an integer N α ≥ 1 N_\alpha\ge1 N α ≥ 1 such that Pr ( S q / q ≥ v / 2 ) ≥ α \Pr(S_q/q\ge v/2)\ge\alpha Pr ( S q / q ≥ v /2 ) ≥ α for every integer q ≥ N α q\ge N_\alpha q ≥ N α . Choose an integer c α , d r i f t ≥ max { 1 , N α , 2 / v } c_{\alpha,\mathrm{drift}}\ge\max\{1,N_\alpha,2/v\} c α , drift ≥ max { 1 , N α , 2/ v } . For T ≥ 1 T\ge1 T ≥ 1 and q = c α , d r i f t T q=c_{\alpha,\mathrm{drift}}T q = c α , drift T , we have q ≥ N α q\ge N_\alpha q ≥ N α and v q / 2 ≥ T vq/2\ge T v q /2 ≥ T . Therefore Theorem 2.8 gives Pr i ( τ T ≤ q ) ≥ Pr ( S q ≥ T ) ≥ α \Pr_i(\tau_T\le q)\ge\Pr(S_q\ge T)\ge\alpha Pr i ( τ T ≤ q ) ≥ Pr ( S q ≥ T ) ≥ α . The choice c α , d r i f t ≥ N α c_{\alpha,\mathrm{drift}}\ge N_\alpha c α , drift ≥ N α covers every small positive target. □
In particular R α ( q ) ≥ ⌊ q / c α , d r i f t ⌋ R_\alpha(q)\ge\lfloor q/c_{\alpha,\mathrm{drift}}\rfloor R α ( q ) ≥ ⌊ q / c α , drift ⌋ for every q ∈ N q\in\mathbb{N} q ∈ N , by monotonicity in the budget.
Theorem 3.3 (Positive-drift equality) If E ∣ K ∣ < ∞ \mathbb{E}|K|<\infty E ∣ K ∣ < ∞ and E K > 0 \mathbb{E}K>0 E K > 0 , then P μ , α ∗ = P μ ∗ = P \mathbf{P}_{\mu,\alpha}^*=\mathbf{P}_\mu^*=\mathbf{P} P μ , α ∗ = P μ ∗ = P for every 0 < α < 1 0<\alpha<1 0 < α < 1 .
Proof. The first positive moment is finite, so Theorem 3.1 gives containment in P . Conversely, take a deterministic polynomial-time decider M M M with integer polynomial time bound H H H . For confidence α \alpha α use Q ( n ) = c α , d r i f t ( H ( n ) + 1 ) Q(n)=c_{\alpha,\mathrm{drift}}(H(n)+1) Q ( n ) = c α , drift ( H ( n ) + 1 ) . The preceding lemma reaches H ( n ) + 1 H(n)+1 H ( n ) + 1 , hence every accepting index at most H ( n ) H(n) H ( n ) , with probability at least α \alpha α . The same machine is sound on every no-instance. This proves equality. □
Zero drift
With zero drift and finite nonzero variance, quadratically many ticks give a fixed positive crossing probability. Repeating blocks raises the confidence; Appendix § 13 .§ 2 calibrates the scale and the alternatives.
Lemma 3.4 (Quadratic crossing blocks) If E K = 0 \mathbb{E}K=0 E K = 0 and 0 < σ 2 = E K 2 < ∞ 0<\sigma^2=\mathbb{E}K^2<\infty 0 < σ 2 = E K 2 < ∞ , there are an integer b ≥ 1 b\ge1 b ≥ 1 and a number p ∈ ( 0 , 1 ) p\in(0,1) p ∈ ( 0 , 1 ) such that for every T ≥ 1 T\ge1 T ≥ 1 , i ≥ 0 i\ge0 i ≥ 0 , and integer j ≥ 1 j\ge1 j ≥ 1 ,
Pr i ( τ T > j b T 2 ) ≤ ( 1 − p ) j . \Pr_i(\tau_T>jbT^2)\le(1-p)^j. i Pr ( τ T > j b T 2 ) ≤ ( 1 − p ) j .
Proof. Let Φ \Phi Φ be the standard normal distribution function and put p = ( 1 − Φ ( 1 ) ) / 2 > 0 p=(1-\Phi(1))/2>0 p = ( 1 − Φ ( 1 )) /2 > 0 . By the classical iid (Lindeberg–Lévy) central limit theorem [18 Theorem 3.4.1] , there is an integer N ≥ 1 N\ge1 N ≥ 1 such that
Pr ( S q σ q ≥ 1 ) ≥ p for every q ≥ N . \Pr\left(\frac{S_q}{\sigma\sqrt{q}}\ge1\right)\ge p\quad\text{for every }q\ge N. Pr ( σ q S q ≥ 1 ) ≥ p for every q ≥ N .
Choose an integer b ≥ max { 1 , N , σ − 2 } b\ge\max\{1,N,\sigma^{-2}\} b ≥ max { 1 , N , σ − 2 } . For q = b T 2 q=bT^2 q = b T 2 , σ q ≥ T \sigma\sqrt{q}\ge T σ q ≥ T and q ≥ N q\ge N q ≥ N , including when T = 1 T=1 T = 1 . Thus Pr ( S q ≥ T ) ≥ p \Pr(S_q\ge T)\ge p Pr ( S q ≥ T ) ≥ p and the free-walk comparison gives Pr i ( τ T ≤ q ) ≥ p \Pr_i(\tau_T\le q)\ge p Pr i ( τ T ≤ q ) ≥ p from every i ≥ 0 i\ge0 i ≥ 0 .
At the end of a block, condition on the complete past and on not yet having hit T T T . The current position is below T T T . The next block uses fresh independent increments, so the same lower bound p p p applies from that position. The tower property [18 Theorem 4.1.13] , applied at these deterministic block boundaries, proves the display. The block-hit events need not be independent. The free-walk bound holds from each starting position, so it supplies the uniform conditional estimate needed here. □
There is also a zero-drift case with infinite variance that follows from a stopped-square argument. Only positive overshoots need to be bounded.
We write E i \mathrm{E}_i E i for expectation under the law of the held walk started at X 0 = i X_0=i X 0 = i , in parallel with the notation Pr i \Pr_i Pr i for probability. For nonnegative quantities a a a and b b b , write a ∧ b = min { a , b } a\wedge b=\min\{a,b\} a ∧ b = min { a , b } . Thus q ∧ τ T q\wedge\tau_T q ∧ τ T is the number of ticks elapsed until either tick q q q or the first crossing of T T T , whichever occurs first, and X q ∧ τ T X_{q\wedge\tau_T} X q ∧ τ T is the clock position at that time.
Lemma 3.5 (positive jumps at zero drift) Suppose E ∣ K ∣ < ∞ \mathrm{E}|K|<\infty E ∣ K ∣ < ∞ , E K = 0 \mathrm{E}K=0 E K = 0 , μ ≠ δ 0 \mu\ne\delta_0 μ = δ 0 , and K ≤ w K\le w K ≤ w almost surely for an integer w ≥ 1 w\ge1 w ≥ 1 . Negative jumps may be unbounded. Then some integer k 0 ≥ 1 k_0\ge1 k 0 ≥ 1 has μ ( k 0 ) > 0 \mu(k_0)>0 μ ( k 0 ) > 0 . Fix one and put δ μ = k 0 2 μ ( k 0 ) > 0 \delta_\mu=k_0^2\mu(k_0)>0 δ μ = k 0 2 μ ( k 0 ) > 0 . Then, for 0 ≤ i < T 0\le i<T 0 ≤ i < T ,
E i τ T ≤ ( T − 1 + w ) 2 − i 2 δ μ , Pr i ( τ T > q ) ≤ ( T − 1 + w ) 2 − i 2 δ μ q ( T , q ≥ 1 ) . \mathrm{E}_i\tau_T\le\frac{(T-1+w)^2-i^2}{\delta_\mu},\qquad
\Pr_i(\tau_T>q)\le\frac{(T-1+w)^2-i^2}{\delta_\mu q}\quad(T,q\ge1). E i τ T ≤ δ μ ( T − 1 + w ) 2 − i 2 , i Pr ( τ T > q ) ≤ δ μ q ( T − 1 + w ) 2 − i 2 ( T , q ≥ 1 ) . Choose an integer b s q ≥ max { 1 , ⌈ 2 w 2 / δ μ ⌉ } b_{\mathrm{sq}}\ge\max\{1,\lceil2w^2/\delta_\mu\rceil\} b sq ≥ max { 1 , ⌈ 2 w 2 / δ μ ⌉} . Then Pr i ( τ T > j b s q T 2 ) ≤ 2 − j \Pr_i(\tau_T>jb_{\mathrm{sq}}T^2)\le2^{-j} Pr i ( τ T > j b sq T 2 ) ≤ 2 − j for j ≥ 1 j\ge1 j ≥ 1 , uniformly below T T T .
Proof. Zero mean and nondegeneracy give a supported k 0 > 0 k_0>0 k 0 > 0 . Set δ μ = μ ( k 0 ) k 0 2 \delta_\mu=\mu(k_0)k_0^2 δ μ = μ ( k 0 ) k 0 2 . For a position i ≥ 0 i\ge0 i ≥ 0 , the conditional squared gain is
E [ X t + 1 2 − X t 2 ∣ X t = i ] = ∑ k ≥ − i ( 2 i k + k 2 ) μ ( k ) ≥ δ μ . \mathrm{E}[X_{t+1}^2-X_t^2\mid X_t=i]=\sum_{k\ge-i}(2ik+k^2)\mu(k)\ge\delta_\mu. E [ X t + 1 2 − X t 2 ∣ X t = i ] = k ≥ − i ∑ ( 2 ik + k 2 ) μ ( k ) ≥ δ μ .
Indeed ∑ k ≥ − i k μ ( k ) = − ∑ k < − i k μ ( k ) ≥ 0 \sum_{k\ge-i}k\mu(k)=-\sum_{k<-i}k\mu(k)\ge0 ∑ k ≥ − i k μ ( k ) = − ∑ k < − i k μ ( k ) ≥ 0 , and the quadratic term includes δ μ \delta_\mu δ μ . These sums are finite because executed jumps lie in [ − i , w ] [-i,w] [ − i , w ] . For τ = τ T \tau=\tau_T τ = τ T , finite summation of conditional gains and the tower property give
E i X q ∧ τ 2 − i 2 ≥ δ μ ∑ s = 0 q − 1 Pr i ( τ > s ) = δ μ E i ( q ∧ τ ) . \mathrm{E}_iX_{q\wedge\tau}^2-i^2\ge\delta_\mu\sum_{s=0}^{q-1}\Pr_i(\tau>s)=\delta_\mu\mathrm{E}_i(q\wedge\tau). E i X q ∧ τ 2 − i 2 ≥ δ μ s = 0 ∑ q − 1 i Pr ( τ > s ) = δ μ E i ( q ∧ τ ) .
The stopped position is at most T − 1 + w T-1+w T − 1 + w . Monotone convergence as q → ∞ q\to\infty q → ∞ gives the expectation bound, and Markov’s inequality gives the tail bound. Since T − 1 + w ≤ T w T-1+w\le Tw T − 1 + w ≤ T w for integers T , w ≥ 1 T,w\ge1 T , w ≥ 1 , the expectation is at most w 2 T 2 / δ μ w^2T^2/\delta_\mu w 2 T 2 / δ μ , uniformly below T T T . Choose an integer b s q ≥ max { 1 , ⌈ 2 w 2 / δ μ ⌉ } b_{\mathrm{sq}}\ge\max\{1,\lceil2w^2/\delta_\mu\rceil\} b sq ≥ max { 1 , ⌈ 2 w 2 / δ μ ⌉} . Markov’s inequality gives success at least 1 / 2 1/2 1/2 in each block of b s q T 2 b_{\mathrm{sq}}T^2 b sq T 2 ticks, uniformly below the target. Fresh conditional blocks iterate as in Theorem 3.4. □
Theorem 3.6 (Zero-drift equality) Suppose E ∣ K ∣ < ∞ \mathrm{E}|K|<\infty E ∣ K ∣ < ∞ , E K = 0 \mathrm{E}K=0 E K = 0 , and μ ≠ δ 0 \mu\ne\delta_0 μ = δ 0 . If either E K 2 < ∞ \mathrm{E}K^2<\infty E K 2 < ∞ or the positive jumps are bounded, then P μ , α ∗ = P μ ∗ = P \mathrm{P}_{\mu,\alpha}^*=\mathrm{P}_\mu^*=\mathrm{P} P μ , α ∗ = P μ ∗ = P for every 0 < α < 1 0<\alpha<1 0 < α < 1 .
Proof. Containment follows from the finite first positive moment. In the finite-variance case, nondegeneracy gives σ 2 > 0 \sigma^2 > 0 σ 2 > 0 . Use Theorem 3.4 with block constant b b b and success bound p p p . In the bounded-positive case use Theorem 3.5 with block constant b s q b_{\mathrm{sq}} b sq and success bound 1 / 2 1/2 1/2 . Denote the chosen constant by b 0 b_0 b 0 and its success bound by p 0 > 0 p_0 > 0 p 0 > 0 . For a polynomial-time decider with bound H H H , choose an integer j α ≥ 1 j_\alpha\ge1 j α ≥ 1 with ( 1 − p 0 ) j α ≤ 1 − α (1-p_0)^{j_\alpha} \le1-\alpha ( 1 − p 0 ) j α ≤ 1 − α . The integer polynomial Q ( n ) = j α b 0 ( H ( n ) + 1 ) 2 Q(n)=j_\alpha b_0(H(n)+1)^2 Q ( n ) = j α b 0 ( H ( n ) + 1 ) 2 is a valid budget. Soundness is unchanged. The excluded law δ 0 \delta_0 δ 0 is covered by Theorem 2.12.
Below P: sequential locality, an intermediate class, and an antichain
Small crossing radii force bounded-prefix recognition through sequential locality. Negative drift gives this class under all positive moments, while sparse laws give intermediate and incomparable classes.
Negative drift gives RDEF
Polynomial Lyapunov bounds give a subpolynomial radius under negative mean and all positive moments [9 Chapters 11 and 14] . When E K − = ∞ \mathbb{E}K^-=\infty E K − = ∞ , the untruncated uniform increment-moment hypothesis of Pemantle–Rosenthal does not apply; their Corollary 3 permits one-sided truncation [11 ] . The proof below uses this principle directly at the level of the held walk’s conditional drift. The class identification then follows from sequential locality.
For an event E E E , 1 E \mathbf{1}_E 1 E is its indicator, equal to one on E E E and zero otherwise; for a random variable Y Y Y , E [ Y ; E ] \mathbb{E}[Y;E] E [ Y ; E ] denotes E [ Y 1 E ] \mathbb{E}[Y\mathbf{1}_E] E [ Y 1 E ] whenever this expectation is defined.
Theorem 4.1 (Negative mean and all positive moments) Suppose Pr ( K > 0 ) > 0 \Pr(K>0)>0 Pr ( K > 0 ) > 0 , E [ ( K + ) s ] < ∞ \mathbb{E}[(K^+)^s]<\infty E [( K + ) s ] < ∞ for every s > 0 s>0 s > 0 , and E K − > E K + \mathbb{E}K^- > \mathbb{E}K^+ E K − > E K + . The last condition permits E K − = ∞ \mathbb{E}K^-=\infty E K − = ∞ , hence extended mean − ∞ -\infty − ∞ . For every integer r ≥ 2 r\ge2 r ≥ 2 there is a finite constant C r C_r C r such that, from zero,
Pr ( τ T ≤ q ) ≤ C r q + 1 ( 1 + T ) r − 1 ( q ∈ N , T ≥ 1 ) . (7) \Pr(\tau_T\le q)\le C_r\frac{q+1}{(1+T)^{r-1}}\qquad(q\in\mathbb{N},\ T\ge1). \tag*{(7)} Pr ( τ T ≤ q ) ≤ C r ( 1 + T ) r − 1 q + 1 ( q ∈ N , T ≥ 1 ) . (7) Consequently, for every 0 < α < 1 0<\alpha<1 0 < α < 1 , P μ , α ∗ = P μ ∗ = R D E F \mathbf{P}^*_{\mu,\alpha}=\mathbf{P}^*_\mu=\mathrm{RDEF} P μ , α ∗ = P μ ∗ = RDEF .
Proof. By monotone convergence [18 Theorem 1.5.7] , choose a finite integer m ≥ 1 m\ge1 m ≥ 1 with
E [ K − ; K − ≤ m ] > E K + . \mathbb{E}[K^-;K^-\le m]>\mathbb{E}K^+. E [ K − ; K − ≤ m ] > E K + .
Replace every proposal K < − m K<-m K < − m by zero, leaving the other proposals unchanged, to obtain a law ν \nu ν . It has integrable strictly negative mean v ν = E K + − E [ K − ; K − ≤ m ] < 0 v_\nu=\mathbb{E}K^+-\mathbb{E}[K^-;K^-\le m]<0 v ν = E K + − E [ K − ; K − ≤ m ] < 0 and the same positive moments. For V r ( i ) = ( 1 + i ) r V_r(i)=(1+i)^r V r ( i ) = ( 1 + i ) r , write Δ μ V r ( i ) \Delta_\mu V_r(i) Δ μ V r ( i ) for its conditional one-step drift at i i i under the held rule. Every removed negative proposal contributes either a decrease of V r V_r V r when executed or zero when held. Thus
Δ μ V r ( i ) ≤ Δ ν V r ( i ) ( i ≥ 0 ) . (8) \Delta_\mu V_r(i)\le\Delta_\nu V_r(i)\qquad(i\ge0). \tag*{(8)} Δ μ V r ( i ) ≤ Δ ν V r ( i ) ( i ≥ 0 ) . (8)
Both drifts are finite at each state: executed negative proposals lie in [ − i , − 1 ] [-i,-1] [ − i , − 1 ] , and positive proposals have finite r r r th moment.
For every fixed integer k k k , as i → ∞ i\to\infty i → ∞ ,
( ( 1 + i + k ) r − ( 1 + i ) r ) 1 { k ≥ − i } ( 1 + i ) r − 1 ⟶ r k . \frac{\bigl((1+i+k)^r-(1+i)^r\bigr)\mathbf{1}_{\{k\ge-i\}}}{(1+i)^{r-1}}\longrightarrow rk. ( 1 + i ) r − 1 ( ( 1 + i + k ) r − ( 1 + i ) r ) 1 { k ≥ − i } ⟶ r k .
For k ≥ 0 k \ge0 k ≥ 0 , expansion bounds the quotient by
∑ a = 1 r ( r a ) k a ( 1 + i ) a − 1 ≤ ( 2 r − 1 ) ( k + k r ) , \sum_{a=1}^{r} \binom{r}{a} \frac{k^a}{(1+i)^{a-1}} \le(2^r-1)(k+k^r), a = 1 ∑ r ( a r ) ( 1 + i ) a − 1 k a ≤ ( 2 r − 1 ) ( k + k r ) ,
using k a ≤ k + k r k^a \le k+k^r k a ≤ k + k r for 1 ≤ a ≤ r 1 \le a \le r 1 ≤ a ≤ r . For k < 0 k<0 k < 0 the mean-value theorem bounds its absolute value by r ∣ k ∣ r|k| r ∣ k ∣ when executed; otherwise it is zero. These bounds are integrable under ν \nu ν , since its negative proposals are bounded and its positive r r r th moment is finite. Dominated convergence [18 Theorem 1.5.8] therefore gives
Δ ν V r ( i ) ( 1 + i ) r − 1 ⟶ r v ν < 0. \frac{\Delta_\nu V_r(i)}{(1+i)^{r-1}} \longrightarrow r v_\nu< 0. ( 1 + i ) r − 1 Δ ν V r ( i ) ⟶ r v ν < 0.
Choose c r > 0 c_r>0 c r > 0 smaller than − r v ν -r v_\nu − r v ν . Beyond a finite state there is then drift at most − c r ( 1 + i ) r − 1 -c_r(1+i)^{r-1} − c r ( 1 + i ) r − 1 . For the finitely many remaining states enlarge a finite B r ≥ 0 B_r \ge0 B r ≥ 0 . By (8), for all i ≥ 0 i \ge0 i ≥ 0 ,
Δ μ V r ( i ) ≤ B r − c r ( 1 + i ) r − 1 . (9) \Delta_\mu V_r(i) \le B_r-c_r(1+i)^{r-1}. \tag*{(9)} Δ μ V r ( i ) ≤ B r − c r ( 1 + i ) r − 1 . (9)
No negative moment of the original law is required. The pathwise bound X t ≤ ∑ s = 1 t K s + X_t \le\sum_{s=1}^{t}K_s^+ X t ≤ ∑ s = 1 t K s + in Theorem 2.8 ensures E V r ( X t ) < ∞ \mathbb{E}V_r(X_t)<\infty E V r ( X t ) < ∞ for every finite t t t . For example, for t ≥ 1 t \ge1 t ≥ 1 the inequality ( ∑ s = 1 t K s + ) r ≤ t r − 1 ∑ s = 1 t ( K s + ) r \left(\sum_{s=1}^{t}K_s^+\right)^r \le t^{r-1}\sum_{s=1}^{t}(K_s^+)^r ( ∑ s = 1 t K s + ) r ≤ t r − 1 ∑ s = 1 t ( K s + ) r proves this. Conditional expectations are thus integrable, and summing (9) through deterministic time q + 1 q+1 q + 1 gives
c r ∑ t = 0 q E ( 1 + X t ) r − 1 ≤ V r ( 0 ) + B r ( q + 1 ) − E V r ( X q + 1 ) ≤ 1 + B r ( q + 1 ) . c_r\sum_{t=0}^{q}\mathbb{E}(1+X_t)^{r-1} \le V_r(0)+B_r(q+1)-\mathbb{E}V_r(X_{q+1}) \le1+B_r(q+1). c r t = 0 ∑ q E ( 1 + X t ) r − 1 ≤ V r ( 0 ) + B r ( q + 1 ) − E V r ( X q + 1 ) ≤ 1 + B r ( q + 1 ) .
A union bound and Markov’s inequality now yield
Pr ( τ T ≤ q ) ≤ ∑ t = 0 q Pr ( X t ≥ T ) ≤ 1 + B r ( q + 1 ) c r ( 1 + T ) r − 1 . \Pr(\tau_T \le q) \le\sum_{t=0}^{q}\Pr(X_t \ge T) \le\frac{1+B_r(q+1)}{c_r(1+T)^{r-1}}. Pr ( τ T ≤ q ) ≤ t = 0 ∑ q Pr ( X t ≥ T ) ≤ c r ( 1 + T ) r − 1 1 + B r ( q + 1 ) .
Taking C r = ( 1 + B r ) / c r C_r=(1+B_r)/c_r C r = ( 1 + B r ) / c r proves (7), including q = 0 q=0 q = 0 . At confidence α \alpha α , whenever R α ( q ) ≥ 1 R_\alpha(q)\ge1 R α ( q ) ≥ 1 the bound gives 1 + R α ( q ) ≤ ( C r ( q + 1 ) / α ) 1 / ( r − 1 ) 1+R_\alpha(q)\le(C_r(q+1)/\alpha)^{1/(r-1)} 1 + R α ( q ) ≤ ( C r ( q + 1 ) / α ) 1/ ( r − 1 ) . The case R α ( q ) = 0 R_\alpha(q)=0 R α ( q ) = 0 is absorbed by a constant. Since r r r can be arbitrarily large, the radius is subpolynomial. Positive supported blocks reach every fixed target by Theorem 2.11. Theorem 2.18(iii) proves the class equality.
Exponential contraction gives a sharper logarithmic radius. The held-boundary correction is handled directly, as in geometric drift estimates [9 , 10 Chapter 15] ; reflected-walk estimates [22 ] do not transfer by Theorem 2.9.
Lemma 4.2 (Exponential bound for the held walk) Suppose g ( κ ) : = E e κ K < 1 g(\kappa):=\mathbb{E}e^{\kappa K}<1 g ( κ ) := E e κ K < 1 for some κ > 0 \kappa>0 κ > 0 . There is a constant C μ ≥ 1 C_\mu\ge1 C μ ≥ 1 , depending only on μ \mu μ and the chosen κ \kappa κ , such that, from zero,
sup t ≥ 0 E e κ X t ≤ C μ , Pr ( τ T ≤ q ) ≤ C μ ( q + 1 ) e − κ T ( T , q ∈ N ) . (10) \sup_{t\ge0}\mathbb{E}e^{\kappa X_t}\le C_\mu,\qquad\Pr(\tau_T\le q)\le C_\mu(q+1)e^{-\kappa T}\quad(T,q\in\mathbb{N}). \tag*{(10)} t ≥ 0 sup E e κ X t ≤ C μ , Pr ( τ T ≤ q ) ≤ C μ ( q + 1 ) e − κ T ( T , q ∈ N ) . (10) Proof. Put V ( i ) = e κ i V(i)=e^{\kappa i} V ( i ) = e κi . The exact conditional identity is
E [ V ( X t + 1 ) ∣ X t = i ] V ( i ) = E [ e κ K ; K ≥ − i ] + Pr ( K < − i ) \frac{\mathbb{E}[V(X_{t+1})\mid X_t=i]}{V(i)}
=\mathbb{E}[e^{\kappa K};K\ge-i]+\Pr(K<-i) V ( i ) E [ V ( X t + 1 ) ∣ X t = i ] = E [ e κ K ; K ≥ − i ] + Pr ( K < − i ) = g ( κ ) + E [ 1 − e κ K ; K < − i ] ≤ g ( κ ) + Pr ( K < − i ) . =g(\kappa)+\mathbb{E}[1-e^{\kappa K};K<-i]\le g(\kappa)+\Pr(K<-i). = g ( κ ) + E [ 1 − e κ K ; K < − i ] ≤ g ( κ ) + Pr ( K < − i ) . Set λ = ( 1 + g ( κ ) ) / 2 < 1 \lambda=(1+g(\kappa))/2<1 λ = ( 1 + g ( κ )) /2 < 1 . Since Pr ( K < − i ) → 0 \Pr(K<-i)\to0 Pr ( K < − i ) → 0 , choose an integer R t a i l ≥ 1 R_{\mathrm{tail}}\ge1 R tail ≥ 1 so that the last expression is at most λ \lambda λ for i ≥ R t a i l i\ge R_{\mathrm{tail}} i ≥ R tail . For the finitely many i < R t a i l i<R_{\mathrm{tail}} i < R tail , the conditional expectation is finite; for example it is at most e κ R t a i l ( g ( κ ) + 1 ) e^{\kappa R_{\mathrm{tail}}}(g(\kappa)+1) e κ R tail ( g ( κ ) + 1 ) . For some finite b b d ≥ 0 b_{\mathrm{bd}}\ge0 b bd ≥ 0 we therefore have
E [ V ( X t + 1 ) ∣ X t = i ] ≤ λ V ( i ) + b b d ( i ≥ 0 ) . \mathbb{E}[V(X_{t+1})\mid X_t=i]\le\lambda V(i)+b_{\mathrm{bd}}\qquad(i\ge0). E [ V ( X t + 1 ) ∣ X t = i ] ≤ λV ( i ) + b bd ( i ≥ 0 ) .
Iterating from V ( 0 ) = 1 V(0)=1 V ( 0 ) = 1 gives E V ( X t ) ≤ λ t + b b d ∑ s = 0 t − 1 λ s ≤ 1 + b b d / ( 1 − λ ) = : C μ \mathbb{E}V(X_t) \le\lambda^t+b_{\mathrm{bd}}\sum_{s=0}^{t-1}\lambda^s \le1+b_{\mathrm{bd}}/(1-\lambda)=:C_\mu E V ( X t ) ≤ λ t + b bd ∑ s = 0 t − 1 λ s ≤ 1 + b bd / ( 1 − λ ) =: C μ . Markov’s inequality gives Pr ( X t ≥ T ) ≤ C μ e − κ T \Pr(X_t\ge T)\le C_\mu e^{-\kappa T} Pr ( X t ≥ T ) ≤ C μ e − κ T . Boole’s inequality (the union bound) over 0 ≤ t ≤ q 0\le t\le q 0 ≤ t ≤ q proves the hitting bound, also at T = 0 T=0 T = 0 and q = 0 q=0 q = 0 .
Corollary 4.3 (Reverse-definite classification) Suppose E e κ K < 1 \mathbb{E}e^{\kappa K}<1 E e κ K < 1 for some κ > 0 \kappa>0 κ > 0 , and Pr ( K > 0 ) > 0 \Pr(K>0)>0 Pr ( K > 0 ) > 0 . For every 0 < α < 1 0<\alpha<1 0 < α < 1 ,
P μ , α ∗ = P μ ∗ = R D E F . \mathsf{P}^{*}_{\mu,\alpha}=\mathsf{P}^{*}_{\mu}=\mathbf{RDEF}. P μ , α ∗ = P μ ∗ = RDEF . Equivalently, a language L L L belongs to this class if and only if some machine accepts exactly $L and has accepting logical time bounded by a constant on every x ∈ L x\in L x ∈ L . Replacing this constant bound by c ( log 2 ( ∣ x ∣ + 1 ) + 1 ) c(\log_2(|x|+1)+1) c ( log 2 ( ∣ x ∣ + 1 ) + 1 ) for a fixed c c c gives the same class.
Proof. The exponential bound makes every positive moment finite, and Jensen’s inequality gives E K < 0 \mathbb{E}K<0 E K < 0 when E K − \mathbb{E}K^{-} E K − is finite; otherwise E K − = ∞ > E K + \mathbb{E}K^{-}=\infty>\mathbb{E}K^{+} E K − = ∞ > E K + . Thus Theorem 4.1 gives the class equality, with the sharper logarithmic radius supplied by Theorem 4.2. Theorem 4.2 bounds the radius by κ − 1 ln ( C μ ( q + 1 ) / α ) \kappa^{-1}\ln(C_\mu(q+1)/\alpha) κ − 1 ln ( C μ ( q + 1 ) / α ) , which is subpolynomial. Positive jump probability gives finite-target budgets by Theorem 2.11. Thus Theorem 2.18(iii) proves the equality. Theorem 2.17 identifies constant accepting time with reverse definiteness. Constant time implies the stated logarithmic bound, and Theorem 2.15 proves the converse. □
With a positive exponential moment, ( g ( κ ) − 1 ) / κ → E K < 0 (g(\kappa)-1)/\kappa\to\mathbb{E}K<0 ( g ( κ ) − 1 ) / κ → E K < 0 as κ ↓ 0 \kappa\downarrow0 κ ↓ 0 , including extended mean − ∞ -\infty − ∞ , by dominated and monotone convergence. If g ( κ 0 ) = 1 g(\kappa_0)=1 g ( κ 0 ) = 1 and K ≢ 0 K\not\equiv0 K ≡ 0 , strict Cauchy–Schwarz gives g ( κ 0 / 2 ) < 1 g(\kappa_0/2)<1 g ( κ 0 /2 ) < 1 and forces positive forward probability; frozen laws are covered by Theorem 2.12. The exponential and non-exponential calibrations are in Appendices E.3 and E.4.
A polynomial tail can overcome negative drift
Drift alone does not classify laws with unbounded support. A sufficiently large positive proposal reaches a target from any nonnegative position, regardless of the earlier negative proposals.
Proposition 4.4 (Polynomial lower bounds on the positive tail) Suppose there are a , c > 0 a,c>0 a , c > 0 such that Pr ( K ≥ T ) ≥ c T − a \Pr(K\ge T)\ge cT^{-a} Pr ( K ≥ T ) ≥ c T − a for all sufficiently large integers T T T . Then P ⊆ P μ , α ∗ \mathbf{P}\subseteq\mathsf{P}^{*}_{\mu,\alpha} P ⊆ P μ , α ∗ for every 0 < α < 1 0<\alpha<1 0 < α < 1 . If also E [ ( K + ) r ] < ∞ \mathbb{E}[(K^{+})^r]<\infty E [( K + ) r ] < ∞ for some r > 0 r>0 r > 0 , equality holds at every confidence and for P μ ∗ \mathsf{P}^{*}_{\mu} P μ ∗ .
Proof. The hypothesis gives unbounded positive jumps, so by decreasing c > 0 c>0 c > 0 we can make the bound hold for every integer T ≥ 1 T\ge1 T ≥ 1 . If any of q q q proposals is at least T T T , the walk hits T T T from any nonnegative start. Independence therefore gives
Pr i ( τ T > q ) ≤ ( 1 − Pr ( K ≥ T ) ) q ≤ e − q c T − a . \Pr_i(\tau_T>q)\le(1-\Pr(K\ge T))^q\le e^{-qcT^{-a}}. i Pr ( τ T > q ) ≤ ( 1 − Pr ( K ≥ T ) ) q ≤ e − q c T − a . The budget ⌈ c − 1 T a ℓ α ⌉ \lceil c^{-1}T^a\ell_\alpha\rceil ⌈ c − 1 T a ℓ α ⌉ suffices. Substituting the logical-time bound of a polynomial-time decider and choosing a dominating integer polynomial proves the inclusion. Target zero requires no ticks. The moment assumption gives the other inclusion by Theorem 3.1 . □
If Pr ( K ≥ T ) \Pr(K\ge T) Pr ( K ≥ T ) is bounded above and below by positive constant multiples of T − a T^{-a} T − a for some fixed a > 0 a>0 a > 0 , the class is P \mathbf{P} P . For 0 < r < min ( a , 1 ) 0<r<\min(a,1) 0 < r < min ( a , 1 ) , the tail-sum identity, an application of Tonelli’s theorem to nonnegative summands [18 Theorem 1.7.2] , is
E [ ( K + ) r ] = ∑ m = 1 ∞ ( m r − ( m − 1 ) r ) Pr ( K ≥ m ) . \mathbb{E}[(K^{+})^r]=\sum_{m=1}^{\infty}\left(m^r-(m-1)^r\right)\Pr(K\ge m). E [( K + ) r ] = m = 1 ∑ ∞ ( m r − ( m − 1 ) r ) Pr ( K ≥ m ) .
Since m r − ( m − 1 ) r = O ( m r − 1 ) m^r-(m-1)^r=O(m^{r-1}) m r − ( m − 1 ) r = O ( m r − 1 ) , the moment is finite. The same conclusion applies to any tail asymptotic to c T − a cT^{-a} c T − a . A power-law positive tail with infinite mean therefore does not by itself produce more than P \mathbf{P} P .
In the following example, ζ ( 2 ) = ∑ m = 1 ∞ m − 2 = π 2 / 6 \zeta(2)=\sum_{m=1}^{\infty}m^{-2}=\pi^2/6 ζ ( 2 ) = ∑ m = 1 ∞ m − 2 = π 2 /6 is the Riemann zeta value at $2.
Example 4.5 (Negative drift with class P) Consider the law
μ ( − 1 ) = 3 4 , μ ( j ) = 1 4 ( 1 j 2 − 1 ( j + 1 ) 2 ) ( j ≥ 1 ) . \mu(-1)=\frac{3}{4},\qquad\mu(j)=\frac{1}{4}\left(\frac{1}{j^2}-\frac{1}{(j+1)^2}\right)\quad(j\ge1). μ ( − 1 ) = 4 3 , μ ( j ) = 4 1 ( j 2 1 − ( j + 1 ) 2 1 ) ( j ≥ 1 ) .
The positive masses telescope to 1 / 4 1/4 1/4 , and for every integer T ≥ 1 T\ge1 T ≥ 1 ,
Pr ( K ≥ T ) = 1 4 T 2 , E K + = ∑ T ≥ 1 Pr ( K ≥ T ) = ζ ( 2 ) 4 . \Pr(K\ge T)=\frac{1}{4T^2},\qquad\mathbb{E}K^+=\sum_{T\ge1}\Pr(K\ge T)=\frac{\zeta(2)}{4}. Pr ( K ≥ T ) = 4 T 2 1 , E K + = T ≥ 1 ∑ Pr ( K ≥ T ) = 4 ζ ( 2 ) .
Thus E ∣ K ∣ = ( 3 + ζ ( 2 ) ) / 4 < ∞ \mathbb{E}|K|=(3+\zeta(2))/4<\infty E ∣ K ∣ = ( 3 + ζ ( 2 )) /4 < ∞ and E K = ( ζ ( 2 ) − 3 ) / 4 < 0 \mathbb{E}K=(\zeta(2)-3)/4<0 E K = ( ζ ( 2 ) − 3 ) /4 < 0 (for example ζ ( 2 ) < 2 \zeta(2)<2 ζ ( 2 ) < 2 by integral comparison). The first positive moment is finite. By Theorem 4.4, P μ , α ∗ = P μ ∗ = P \mathrm{P}^*_{\mu,\alpha}=\mathrm{P}^*_\mu=\mathrm{P} P μ , α ∗ = P μ ∗ = P at every confidence. This law has no finite second positive moment. Negative mean alone therefore cannot replace the tail hypotheses in either Theorem 4.3 or Theorem 4.1 . Reversing the roles in the first mixture of Theorem E.2 gives another computable zero-mean law, K = W K=W K = W with probability 1 / ( 1 + m ) 1/(1+m) 1/ ( 1 + m ) and K = − 1 K=-1 K = − 1 otherwise: it has infinite variance and unbounded positive jumps, yet its tail Pr ( K ≥ T ) = T − 2 / ( 1 + m ) \Pr(K\ge T)=T^{-2}/(1+m) Pr ( K ≥ T ) = T − 2 / ( 1 + m ) and finite first moment give class P \mathrm{P} P by Theorem 4.4. Together with Theorem 4.6 , this shows that the zero-drift alternatives are sufficient conditions, not a necessary characterization.
See Appendix § 13 .§ 5 for the order of the one-jump budget.
A zero-drift class strictly between RDEF and P
A sparse zero-drift law lies strictly between bounded-prefix recognition and polynomial time. Its accepting-time radius oscillates between growth scales, and the proof is in Appendix § 9 .
Fix a finite nonempty alphabet Σ \Sigma Σ and let L e v e n = { x ∈ Σ ∗ : ∣ x ∣ is even } L_{\mathrm{even}}=\{x\in\Sigma^*:|x|\text{ is even}\} L even = { x ∈ Σ ∗ : ∣ x ∣ is even } . The next theorem shows that zero drift alone does not give P \mathrm{P} P .
Theorem 4.6 (A strictly intermediate zero-drift class) There is one computable law μ \mu μ , independent of the finite nonempty input alphabet Σ \Sigma Σ , supported on { − 1 , 0 } ∪ { N j : j ≥ 1 } \{-1,0\}\cup\{N_j:j\ge1\} { − 1 , 0 } ∪ { N j : j ≥ 1 } , for strictly increasing positive integers N j N_j N j , such that
E ∣ K ∣ = 1 , E K = 0 , E K 2 = ∞ , L e v e n ∉ P μ , α ∗ ( 0 < α < 1 ) . \mathbb{E}|K|=1,\qquad\mathbb{E}K=0,\qquad\mathbb{E}K^2=\infty,\qquad L_{\mathrm{even}}\notin\mathrm{P}^*_{\mu,\alpha}\quad(0<\alpha<1). E ∣ K ∣ = 1 , E K = 0 , E K 2 = ∞ , L even ∈ / P μ , α ∗ ( 0 < α < 1 ) . For every finite nonempty Σ \Sigma Σ and every 0 < α < 1 0<\alpha<1 0 < α < 1 , R D E F ⊊ P μ , α ∗ ⊊ P \mathrm{RDEF}\subsetneq\mathrm{P}^*_{\mu,\alpha}\subsetneq\mathrm{P} RDEF ⊊ P μ , α ∗ ⊊ P , and likewise R D E F ⊊ P μ ∗ ⊊ P \mathrm{RDEF}\subsetneq\mathrm{P}^*_\mu\subsetneq\mathrm{P} RDEF ⊊ P μ ∗ ⊊ P .
Remark 4.7 (Oscillating radius) The radius of this law oscillates between growth scales. At q j = T j q_j=T_j q j = T j , (A5) gives R α ( q j ) < T j = q j 1 / j R_\alpha(q_j)<T_j=q_j^{1/j} R α ( q j ) < T j = q j 1/ j for large j j j . Along q = ⌈ 2 ℓ α N j 2 / 3 ⌉ q=\lceil2\ell_\alpha N_j^{2/3}\rceil q = ⌈ 2 ℓ α N j 2/3 ⌉ , (A6) gives R α ( q ) ≥ c α q 2 / 3 R_\alpha(q)\ge c_\alpha q^{2/3} R α ( q ) ≥ c α q 2/3 for a constant c α > 0 c_\alpha>0 c α > 0 . Thus neither the eventual positive-power lower bound nor the subpolynomial upper bound in Theorem 2.18 holds. Its polynomial upper-bound criterion still applies, but none of its exact-class criteria does; the exact class remains open.
Infinitely many incomparable classes
Theorem 4.6 gives one class strictly between R D E F \mathrm{RDEF} RDEF and P \mathrm{P} P . Sparse negative-drift laws give infinitely many, pairwise incomparable.
Let t w 1 = 32 \mathrm{tw}_1=32 tw 1 = 32 and t w j + 1 = 2 t w j \mathrm{tw}_{j+1}=2^{\mathrm{tw}_j} tw j + 1 = 2 tw j . For e ∈ N e\in\mathbb{N} e ∈ N let J e = { 2 e ( 2 m + 1 ) : m ∈ N } J_e=\{2^{e}(2m+1):m\in\mathbb{N}\} J e = { 2 e ( 2 m + 1 ) : m ∈ N } ; these sets partition { 1 , 2 , … } \{1,2,\ldots\} { 1 , 2 , … } . Define the law μ ⟨ e ⟩ \mu^{\langle e\rangle} μ ⟨ e ⟩ by
μ ⟨ e ⟩ ( − 1 ) = 1 2 , μ ⟨ e ⟩ ( t w j ) = t w j − 4 ( j ∈ J e ) , \mu^{\langle e\rangle}(-1)=\frac{1}{2},\qquad\mu^{\langle e\rangle}(\mathrm{tw}_j)=\mathrm{tw}_j^{-4}\quad(j\in J_e), μ ⟨ e ⟩ ( − 1 ) = 2 1 , μ ⟨ e ⟩ ( tw j ) = tw j − 4 ( j ∈ J e ) ,
μ ⟨ e ⟩ ( 0 ) = 1 2 − ∑ j ∈ J e t w j − 4 , (11) \mu^{\langle e\rangle}(0)=\frac{1}{2}-\sum_{j\in J_e}\mathrm{tw}_j^{-4}, \tag*{(11)} μ ⟨ e ⟩ ( 0 ) = 2 1 − j ∈ J e ∑ tw j − 4 , (11)
with zero mass elsewhere. For an integer j 0 ≥ 1 j_0 \ge1 j 0 ≥ 1 let
L j 0 ( e ) = { x ∈ Σ ∗ : ∣ x ∣ = tw j for some j ∈ J e with j ≥ j 0 } . L^{(e)}_{j_0}=\{x\in\Sigma^*:|x|=\operatorname{tw}_j\text{ for some }j\in J_e\text{ with }j\ge j_0\}. L j 0 ( e ) = { x ∈ Σ ∗ : ∣ x ∣ = tw j for some j ∈ J e with j ≥ j 0 } .
The sparse laws separate from one another at every pair of confidences.
Theorem 4.8 (An antichain of classes below P) Each μ ( e ) \mu^{(e)} μ ( e ) is computable, has EK < 0 \operatorname{EK}<0 EK < 0 , and, for s > 0 s>0 s > 0 , has E [ ( K + ) s ] < ∞ \mathbb{E}[(K^+)^s]<\infty E [( K + ) s ] < ∞ exactly when s < 4 / 3 s<4/3 s < 4/3 . Fix a finite nonempty alphabet Σ \Sigma Σ . There is j ∗ j_* j ∗ such that for every j 0 ≥ j ∗ j_0\ge j_* j 0 ≥ j ∗ , every e ≠ e ′ e\ne e' e = e ′ , and all α , α ′ ∈ ( 0 , 1 ) \alpha,\alpha'\in(0,1) α , α ′ ∈ ( 0 , 1 ) *,
L j 0 ( e ) ∈ P μ ( e ) , α ∗ ⊆ P μ ( e ′ ) , α ′ ∗ L^{(e)}_{j_0}\in\mathbf{P}^*_{\mu^{(e)},\alpha}\subseteq\mathbf{P}^*_{\mu^{(e')},\alpha'} L j 0 ( e ) ∈ P μ ( e ) , α ∗ ⊆ P μ ( e ′ ) , α ′ ∗ L j 0 ( e ) ∉ P μ ( e ′ ) , α ′ ∗ . L^{(e)}_{j_0}\notin\mathbf{P}^*_{\mu^{(e')},\alpha'}. L j 0 ( e ) ∈ / P μ ( e ′ ) , α ′ ∗ . Consequently, for any confidences, the classes P μ ( e ) , α ∗ \mathbf{P}^*_{\mu^{(e)},\alpha} P μ ( e ) , α ∗ , e ∈ N e\in\mathbb{N} e ∈ N , are pairwise incomparable, and each satisfies
RDEF ⊈ P μ ( e ) , α ∗ ⊊ P . \operatorname{RDEF}\not\subseteq\mathbf{P}^*_{\mu^{(e)},\alpha}\subsetneq\mathbf{P}. RDEF ⊆ P μ ( e ) , α ∗ ⊊ P . The same holds for the classes P μ ( e ) ∗ \mathbf{P}^*_{\mu^{(e)}} P μ ( e ) ∗ .
These laws have negative mean and finite positive moments of every order below 4 / 3 4/3 4/3 , a regime that Theorem 4.1 (all positive moments finite, class RDEF ) and Theorem 4.5 (class P ) leave open. Fixed-law classes below P therefore do not form a chain. The proof is in Appendix § 10 .
Classes below P need negative proposals
Nontrivial iid classes below P require negative proposals. The deletion comparison explains their role.
Above P: the tail determines the class
Tail quantiles organise the iid classes above P . We first give the tail sandwich, then construct exponential, quasipolynomial and doubly exponential classes and identify the effectiveness boundary.
Tail sandwich
For a nondecreasing function f : N → N f:\mathbb{N}\to\mathbb{N} f : N → N let T ( f ) \mathcal{T}(f) T ( f ) be the set of languages L L L for which some deterministic machine M M M accepts exactly L L L and some allowed polynomial Q Q Q satisfies A M ( x ) ≤ f ( Q ( ∣ x ∣ ) ) A_M(x)\le f(Q(|x|)) A M ( x ) ≤ f ( Q ( ∣ x ∣ )) for every x ∈ L x\in L x ∈ L . By Theorem 2.4, P μ , α ∗ = T ( R α ) \mathbf{P}^*_{\mu,\alpha}=\mathcal{T}(R_\alpha) P μ , α ∗ = T ( R α ) .
A proposal-tail quantile controls the class up to the multiplicative tick budget.
Proposition 5.1 (Tail sandwich) For every iid law μ \mu μ and 0 < α < 1 0<\alpha<1 0 < α < 1 , with u α u_\alpha u α as in Theorem 2.10,
T ( u α ) ⊆ P μ , α ∗ ⊆ T ( q ↦ q u α ( q ) ) . \mathcal{T}(u_\alpha)\subseteq\mathbf{P}^*_{\mu,\alpha}\subseteq\mathcal{T}(q\mapsto q\,u_\alpha(q)). T ( u α ) ⊆ P μ , α ∗ ⊆ T ( q ↦ q u α ( q )) .
Proof. Theorem 2.10 gives u α ( q ) ≤ R α ( q ) ≤ q u α ( q ) u_\alpha(q)\le R_\alpha(q)\le q\,u_\alpha(q) u α ( q ) ≤ R α ( q ) ≤ q u α ( q ) for q ≥ 1 q\ge1 q ≥ 1 , and all three vanish at q = 0 q=0 q = 0 . The function u α u_\alpha u α is nondecreasing in q q q , because Pr ( J q ≥ u ) \Pr(J_q\ge u) Pr ( J q ≥ u ) is nondecreasing in q q q . Apply T \mathcal{T} T , which is monotone in f f f .
Thus the class of an iid law is determined, up to the factor q q q , by the quantiles of a single proposal tail. The next laws exploit this.
One law with class EXP
A tail of order 1 / log T 1/\log T 1/ log T crosses exponentially distant indices with a polynomial budget. The same tail bounds the class from above.
Definition 5.2 (The exponential-time law) Let μ E ( 2 j ) = 1 / ( j ( j + 1 ) ) \mu_{\mathrm{E}}(2^j)=1/(j(j+1)) μ E ( 2 j ) = 1/ ( j ( j + 1 )) for integers j ≥ 1 j\ge1 j ≥ 1 , with zero mass elsewhere. For a real z > 0 z>0 z > 0 , write h ( z ) = max { 1 , ⌈ log 2 z ⌉ } h(z)=\max\{1,\lceil\log_2 z\rceil\} h ( z ) = max { 1 , ⌈ log 2 z ⌉} .
The law is normalized because ∑ j = 1 N 1 / ( j ( j + 1 ) ) = 1 − 1 / ( N + 1 ) \sum_{j=1}^{N}1/(j(j+1))=1-1/(N+1) ∑ j = 1 N 1/ ( j ( j + 1 )) = 1 − 1/ ( N + 1 ) . Its exact tail, including 0 < z ≤ 2 0<z\le2 0 < z ≤ 2 , is
Pr ( K ≥ z ) = ∑ j ≥ h ( z ) 1 j ( j + 1 ) = 1 h ( z ) . (12) \Pr(K\ge z)=\sum_{j\ge h(z)}\frac{1}{j(j+1)}=\frac{1}{h(z)}. \tag*{(12)} Pr ( K ≥ z ) = j ≥ h ( z ) ∑ j ( j + 1 ) 1 = h ( z ) 1 . (12)
The ceiling matters at powers of two. Every positive moment is infinite, since 2 r j / ( j ( j + 1 ) ) 2^{rj}/(j(j+1)) 2 r j / ( j ( j + 1 )) does not tend to zero for r > 0 r>0 r > 0 . The extended mean is + ∞ +\infty + ∞ ; there is no finite drift.
Lemma 5.3 (Two bounds for the exponential-time law) For μ = μ E \mu=\mu_{\mathrm{E}} μ = μ E , T ≥ 1 T\ge1 T ≥ 1 , and integer q ≥ 1 q\ge1 q ≥ 1 ,
Pr ( τ T > q ) ≤ e − q / h ( T ) , Pr ( τ T ≤ q ) ≤ q h ( T / q ) . (13) \Pr(\tau_T>q)\le e^{-q/h(T)},\qquad\Pr(\tau_T\le q)\le\frac{q}{h(T/q)}. \tag*{(13)} Pr ( τ T > q ) ≤ e − q / h ( T ) , Pr ( τ T ≤ q ) ≤ h ( T / q ) q . (13) If Pr ( τ T ≤ q ) ≥ α > 0 \Pr(\tau_T\le q)\ge\alpha>0 Pr ( τ T ≤ q ) ≥ α > 0 , then
T ≤ q 2 q / α . (14) T\le q2^{q/\alpha}. \tag*{(14)} T ≤ q 2 q / α . (14) For q = 0 q=0 q = 0 , positive confidence is possible only at T = 0 T=0 T = 0 .
Proof. Apply Theorem 2.10 with the tail (12) . The first inclusion and 1 − u ≤ e − u 1-u\le e^{-u} 1 − u ≤ e − u give the failure bound; the second inclusion and a union bound give the success bound. At positive confidence the latter implies h ( T / q ) ≤ q / α h(T/q)\le q/\alpha h ( T / q ) ≤ q / α . Since log 2 ( T / q ) ≤ h ( T / q ) \log_2(T/q)\le h(T/q) log 2 ( T / q ) ≤ h ( T / q ) for every positive T / q T/q T / q , exponentiation gives (14) .
For a polynomial budget Q Q Q , (14) bounds each accepting index by Q ( n ) 2 Q ( n ) / α Q(n)\,2^{Q(n)/\alpha} Q ( n ) 2 Q ( n ) / α . Conversely, the failure bound gives exponential reachability: if k α ≥ max { 1 , ℓ α } k_\alpha\ge\max\{1,\ell_\alpha\} k α ≥ max { 1 , ℓ α } is an integer, then R α ( q ) ≥ 2 ⌊ q / k α ⌋ R_\alpha(q)\ge2^{\lfloor q/k_\alpha\rfloor} R α ( q ) ≥ 2 ⌊ q / k α ⌋ for q ≥ k α q\ge k_\alpha q ≥ k α , while (14) gives R α ( q ) ≤ q 2 q / α R_\alpha(q)\le q2^{q/\alpha} R α ( q ) ≤ q 2 q / α for q ≥ 1 q\ge1 q ≥ 1 .
Proposition 5.4 (Sharp fixed-confidence scale for μ E \mu_{\mathrm{E}} μ E ) For T , q ≥ 1 T,q\ge1 T , q ≥ 1 ,
1 − ( 1 − 1 h ( T ) ) q ≤ Pr ( τ T ≤ q ) ≤ 1 − ( 1 − 1 h ( T / q ) ) q . (15) 1-\left(1-\frac{1}{h(T)}\right)^q\le\Pr(\tau_T\le q)\le1-\left(1-\frac{1}{h(T/q)}\right)^q. \tag*{(15)} 1 − ( 1 − h ( T ) 1 ) q ≤ Pr ( τ T ≤ q ) ≤ 1 − ( 1 − h ( T / q ) 1 ) q . (15) For every fixed c > 0 c>0 c > 0 and 0 < α < 1 0<\alpha<1 0 < α < 1 , as T → ∞ T\to\infty T → ∞ ,
Pr ( τ T ≤ ⌊ c log 2 T ⌋ ) ⟶ 1 − e − c , q α ( T ) log 2 T ⟶ ℓ α . \Pr(\tau_T\le\lfloor c\log_2 T\rfloor)\longrightarrow1-e^{-c},\qquad\frac{q_\alpha(T)}{\log_2 T}\longrightarrow\ell_\alpha. Pr ( τ T ≤ ⌊ c log 2 T ⌋) ⟶ 1 − e − c , log 2 T q α ( T ) ⟶ ℓ α .
Proof. Theorem 2.10 and (12) give the sandwich, including T ≤ 2 T\le2 T ≤ 2 and T / q ≤ 2 T/q\le2 T / q ≤ 2 . For ℓ T = log 2 T \ell_T=\log_2T ℓ T = log 2 T and q = ⌊ c ℓ T ⌋ q=\lfloor c\ell_T\rfloor q = ⌊ c ℓ T ⌋ , we have h ( T ) = ℓ T + O ( 1 ) h(T)=\ell_T+O(1) h ( T ) = ℓ T + O ( 1 ) and h ( T / q ) = ℓ T − log 2 q + O ( 1 ) h(T/q)=\ell_T-\log_2q+O(1) h ( T / q ) = ℓ T − log 2 q + O ( 1 ) . Since log 2 q = O ( log ℓ T ) \log_2q=O(\log\ell_T) log 2 q = O ( log ℓ T ) , both h ( T ) / ℓ T h(T)/\ell_T h ( T ) / ℓ T and h ( T / q ) / ℓ T h(T/q)/\ell_T h ( T / q ) / ℓ T tend to 1 1 1 . Thus both failure probabilities tend to e − c e^{-c} e − c . Fix 0 < ε < ℓ α 0<\varepsilon<\ell_\alpha 0 < ε < ℓ α . The limits at ℓ α − ε \ell_\alpha-\varepsilon ℓ α − ε and ℓ α + ε \ell_\alpha+\varepsilon ℓ α + ε put q α ( T ) q_\alpha(T) q α ( T ) between the corresponding integer budgets for all sufficiently large T T T ; divide by ℓ T \ell_T ℓ T and let ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 .
The logarithmic crossing budget gives exactly exponential-time recognition.
Theorem 5.5 (Exact exponential-time class) For every 0 < α < 1 0<\alpha<1 0 < α < 1 ,
P μ E , α ∗ = P μ E ∗ = E X P . \mathsf{P}^{*}_{\mu_{\mathrm{E}},\alpha}=\mathsf{P}^{*}_{\mu_{\mathrm{E}}}=\mathrm{EXP}. P μ E , α ∗ = P μ E ∗ = EXP . Proof. The bounds immediately preceding Theorem 5.4 give an eventual exponential lower radius bound and R α ( q ) ≤ q 2 q / α R_\alpha(q)\le q2^{q/\alpha} R α ( q ) ≤ q 2 q / α for q ≥ 1 q\ge1 q ≥ 1 . The latter is bounded by 2 C ( q + 1 ) 2^{C(q+1)} 2 C ( q + 1 ) for a fixed C C C . Apply Theorem 2.18(iv), separately at each confidence.
The law μ E \mu_{\mathrm{E}} μ E has computable rational masses and an effective tail. Exact sampling and its infinite expected writing cost are discussed in Appendix § 13 .§ 6 .
The hierarchy QP and 2EXP
Varying one tail produces infinitely many fixed-law classes. A further computable tail reaches doubly exponential time.
For functions of the target T T T , write f ( T ) ≍ g ( T ) f(T) \asymp g(T) f ( T ) ≍ g ( T ) as T → ∞ T \to\infty T → ∞ if c g ( T ) ≤ f ( T ) ≤ C g ( T ) c g(T) \le f(T) \le C g(T) c g ( T ) ≤ f ( T ) ≤ C g ( T ) for some constants 0 < c ≤ C < ∞ 0 < c \le C < \infty 0 < c ≤ C < ∞ and all sufficiently large T T T . The constants are independent of T T T and may depend on fixed parameters unless uniformity is expressly asserted.
Definition 5.6 (Hierarchy laws) For an integer k ≥ 1 k \ge1 k ≥ 1 let J J J satisfy Pr ( J = j ) = 2 − j \Pr(J = j) = 2^{-j} Pr ( J = j ) = 2 − j for j ≥ 1 j \ge1 j ≥ 1 , and let μ [ k ] \mu_{[k]} μ [ k ] be the law of K = 2 J k K = 2^{J^k} K = 2 J k . Put h k ( T ) = 1 h_k(T) = 1 h k ( T ) = 1 for T ≤ 2 T \le2 T ≤ 2 and h k ( T ) = ⌈ ( log 2 T ) 1 / k ⌉ h_k(T) = \lceil(\log_2 T)^{1/k}\rceil h k ( T ) = ⌈( log 2 T ) 1/ k ⌉ for T > 2 T > 2 T > 2 . Then Pr ( K ≥ T ) = Pr ( J ≥ h k ( T ) ) = 2 1 − h k ( T ) \Pr(K \ge T) = \Pr(J \ge h_k(T)) = 2^{1-h_k(T)} Pr ( K ≥ T ) = Pr ( J ≥ h k ( T )) = 2 1 − h k ( T ) , and for every T ≥ 1 T \ge1 T ≥ 1
2 − ( log 2 T ) 1 / k ≤ Pr ( K ≥ T ) ≤ 2 1 − ( log 2 T ) 1 / k . (16) 2^{-(\log_2 T)^{1/k}} \le\Pr(K \ge T) \le2^{1-(\log_2 T)^{1/k}}. \tag*{(16)} 2 − ( l o g 2 T ) 1/ k ≤ Pr ( K ≥ T ) ≤ 2 1 − ( l o g 2 T ) 1/ k . (16)
The point masses are computable rational numbers, so μ [ k ] \mu_{[k]} μ [ k ] is computable. The law μ [ 1 ] \mu_{[1]} μ [ 1 ] has Pr ( K ≥ T ) ≍ T − 1 \Pr(K \ge T) \asymp T^{-1} Pr ( K ≥ T ) ≍ T − 1 and class P \mathbf{P} P by Theorem 4.4; its positive moments of orders 0 < r < 1 0 < r < 1 0 < r < 1 are finite.
Each hierarchy law gives exactly one quasipolynomial-time level, independently of confidence.
Theorem 5.7 (A hierarchy of fixed-law classes) For every integer k ≥ 1 k \ge1 k ≥ 1 and every 0 < α < 1 0 < \alpha< 1 0 < α < 1 ,
P μ [ k ] , α ∗ = P μ [ k ] ∗ = Q P k . \mathbf{P}^{*}_{\mu_{[k]},\alpha} = \mathbf{P}^{*}_{\mu_{[k]}} = \mathbf{QP}_k. P μ [ k ] , α ∗ = P μ [ k ] ∗ = QP k . Consequently P = Q P 1 ⊊ Q P 2 ⊊ ⋯ ⊊ E X P \mathbf{P} = \mathbf{QP}_1 \subsetneq\mathbf{QP}_2 \subsetneq\cdots\subsetneq\mathbf{EXP} P = QP 1 ⊊ QP 2 ⊊ ⋯ ⊊ EXP : computable iid laws realise infinitely many distinct fixed classes between P \mathbf{P} P and E X P \mathbf{EXP} EXP , none of which depends on the confidence.
Proof. Quantile bounds. By Theorem 2.10, an integer u ≥ 1 u \ge1 u ≥ 1 is eligible for u α ( q ) u_\alpha(q) u α ( q ) exactly when Pr ( K ≥ u ) ≥ p q : = 1 − ( 1 − α ) 1 / q \Pr(K \ge u) \ge p_q := 1 - (1-\alpha)^{1/q} Pr ( K ≥ u ) ≥ p q := 1 − ( 1 − α ) 1/ q . Bernoulli’s inequality and 1 − e − x ≤ x 1-e^{-x} \le x 1 − e − x ≤ x give α / q ≤ p q ≤ ℓ α / q \alpha/q \le p_q \le\ell_\alpha/q α / q ≤ p q ≤ ℓ α / q . If q ≥ ℓ α q \ge\ell_\alpha q ≥ ℓ α , the integer u = ⌊ 2 ( log 2 ( q / ℓ α ) ) k ⌋ ≥ 1 u = \left\lfloor2^{(\log_2(q/\ell_\alpha))^k}\right\rfloor\ge1 u = ⌊ 2 ( l o g 2 ( q / ℓ α ) ) k ⌋ ≥ 1 satisfies ( log 2 u ) 1 / k ≤ log 2 ( q / ℓ α ) (\log_2 u)^{1/k} \le\log_2(q/\ell_\alpha) ( log 2 u ) 1/ k ≤ log 2 ( q / ℓ α ) , hence Pr ( K ≥ u ) ≥ ℓ α / q ≥ p q \Pr(K \ge u) \ge\ell_\alpha/q \ge p_q Pr ( K ≥ u ) ≥ ℓ α / q ≥ p q by (15). Conversely, an eligible u ≥ 1 u \ge1 u ≥ 1 satisfies 2 1 − ( log 2 u ) 1 / k ≥ α / q 2^{1-(\log_2 u)^{1/k}} \ge\alpha/q 2 1 − ( l o g 2 u ) 1/ k ≥ α / q , so log 2 u ≤ ( log 2 ( 2 q / α ) ) k \log_2 u \le(\log_2(2q/\alpha))^k log 2 u ≤ ( log 2 ( 2 q / α ) ) k . Hence
⌊ 2 ( log 2 ( q / ℓ α ) ) k ⌋ ≤ R α ( q ) ≤ q 2 ( log 2 ( 2 q / α ) ) k ( q ≥ ℓ α , q ≥ 1 ) . (17) \left\lfloor2^{(\log_2(q/\ell_\alpha))^k}\right\rfloor\le R_\alpha(q) \le q\,2^{(\log_2(2q/\alpha))^k}\qquad(q \ge\ell_\alpha,\ q \ge1). \tag*{(17)} ⌊ 2 ( l o g 2 ( q / ℓ α ) ) k ⌋ ≤ R α ( q ) ≤ q 2 ( l o g 2 ( 2 q / α ) ) k ( q ≥ ℓ α , q ≥ 1 ) . (17)
The upper bound holds for every q ≥ 1 q \ge1 q ≥ 1 .
Upper inclusion. Let ( M , Q ) (M,Q) ( M , Q ) be a witness and choose integers a , d ≥ 1 a,d \ge1 a , d ≥ 1 with Q ( n ) ≤ a ( n + 2 ) d Q(n) \le a(n+2)^d Q ( n ) ≤ a ( n + 2 ) d . Put c 1 = d + max { 0 , log 2 ( 2 a / α ) } c_1 = d + \max\{0,\log_2(2a/\alpha)\} c 1 = d + max { 0 , log 2 ( 2 a / α )} . Since log 2 ( n + 2 ) ≥ 1 \log_2(n+2) \ge1 log 2 ( n + 2 ) ≥ 1 , the upper bound gives, whenever Q ( n ) ≥ 1 Q(n) \ge1 Q ( n ) ≥ 1 ,
A M ( x ) ≤ Q ( n ) 2 ( log 2 ( 2 Q ( n ) / α ) ) k A_M(x) \le Q(n)\,2^{(\log_2(2Q(n)/\alpha))^k} A M ( x ) ≤ Q ( n ) 2 ( l o g 2 ( 2 Q ( n ) / α ) ) k
≤ 2 c 1 log 2 ( n + 2 ) 2 ( c 1 log 2 ( n + 2 ) ) k \le2^{c_1\log_2(n+2)}\,2^{(c_1\log_2(n+2))^k} ≤ 2 c 1 l o g 2 ( n + 2 ) 2 ( c 1 l o g 2 ( n + 2 ) ) k
≤ 2 2 c 1 k ( log 2 ( n + 2 ) ) k \le2^{2c_1^k(\log_2(n+2))^k} ≤ 2 2 c 1 k ( l o g 2 ( n + 2 ) ) k
for x ∈ L x \in L x ∈ L of length n n n ; when Q ( n ) = 0 Q(n) = 0 Q ( n ) = 0 , A M ( x ) = 0 A_M(x) = 0 A M ( x ) = 0 . A decider computes ⌈ log 2 ( n + 2 ) ⌉ \lceil\log_2(n+2)\rceil ⌈ log 2 ( n + 2 )⌉ , which is at most 2 log 2 ( n + 2 ) 2\log_2(n+2) 2 log 2 ( n + 2 ) , and simulates M M M through 2 c 2 ⌈ log 2 ( n + 2 ) ⌉ k 2^{c_2\lceil\log_2(n+2)\rceil^k} 2 c 2 ⌈ l o g 2 ( n + 2 ) ⌉ k steps for a fixed integer c 2 ≥ 2 c 1 k c_2 \ge2c_1^k c 2 ≥ 2 c 1 k . It accepts exactly when that simulation accepts, including at index zero. Soundness makes it a decider for L L L , and its running time on the model of Theorem 2.1 is polynomial in that bound. Hence L ∈ Q P k L \in\mathbf{QP}_k L ∈ QP k .
Lower inclusion. Let a decider run within 2 c ( log 2 ( n + 2 ) ) k 2^{c(\log_2(n+2))^k} 2 c ( l o g 2 ( n + 2 ) ) k transitions, with c ≥ 1 c \ge1 c ≥ 1 an integer. Put Q ( n ) = ⌈ ℓ α ⌉ ( n + 2 ) c Q(n) = \lceil\ell_\alpha\rceil(n+2)^c Q ( n ) = ⌈ ℓ α ⌉ ( n + 2 ) c , an allowed polynomial. Then Q ( n ) ≥ ℓ α Q(n) \ge\ell_\alpha Q ( n ) ≥ ℓ α and log 2 ( Q ( n ) / ℓ α ) ≥ c log 2 ( n + 2 ) \log_2(Q(n)/\ell_\alpha) \ge c\log_2(n+2) log 2 ( Q ( n ) / ℓ α ) ≥ c log 2 ( n + 2 ) , so (16) gives
R α ( Q ( n ) ) ≥ ⌊ 2 c k ( log 2 ( n + 2 ) ) k ⌋ ≥ ⌊ 2 c ( log 2 ( n + 2 ) ) k ⌋ , \begin{aligned}
R_\alpha(Q(n)) \ge\left\lfloor2^{c^k(\log_2(n+2))^k}\right\rfloor\\
&\ge\left\lfloor2^{c(\log_2(n+2))^k}\right\rfloor,
\end{aligned} R α ( Q ( n )) ≥ ⌊ 2 c k ( l o g 2 ( n + 2 ) ) k ⌋ ≥ ⌊ 2 c ( l o g 2 ( n + 2 ) ) k ⌋ ,
e n d s m i d = 1 {ends_mid=1} e n d s m i d = 1
which bounds every accepting index. Theorem 2.4 gives L ∈ P μ [ k ] , α ∗ L \in\mathrm{P}^{*}_{\mu[k],\alpha} L ∈ P μ [ k ] , α ∗ , with a witness depending on α \alpha α only through Q Q Q .
Strictness. Use integer time bounds
B ( n ) = ⌈ log 2 ( n + 2 ) ⌉ , f k ( n ) = 2 B ( n ) k B ( n ) . B(n)=\left\lceil\log_{2}(n+2)\right\rceil,\qquad f_k(n)=2^{B(n)^k}\sqrt{B(n)}. B ( n ) = ⌈ log 2 ( n + 2 ) ⌉ , f k ( n ) = 2 B ( n ) k B ( n ) .
These functions and their squares are time-constructible on multitape machines: their integer exponents are computable in polynomial time in n + 2 n+2 n + 2 , and a binary counter supplies a clock in O ( f k ( n ) ) O(f_k(n)) O ( f k ( n )) time. Every language in Q P k \mathrm{QP}_k QP k lies in D T I M E ( f k ) \mathrm{DTIME}(f_k) DTIME ( f k ) , also on multitape machines. Since f k log f k = o ( f k 2 ) f_k\log f_k=o(f_k^2) f k log f k = o ( f k 2 ) , the multitape time hierarchy theorem [16 Section 3.1] gives a language in D T I M E ( f k 2 ) ∖ D T I M E ( f k ) \mathrm{DTIME}(f_k^2)\setminus\mathrm{DTIME}(f_k) DTIME ( f k 2 ) ∖ DTIME ( f k ) . The quadratic single-tape simulation places it in D T I M E ( f k 4 ) ⊆ Q P k + 1 \mathrm{DTIME}(f_k^4)\subseteq\mathrm{QP}_{k+1} DTIME ( f k 4 ) ⊆ QP k + 1 , since log 2 f k ( n ) = Θ ( ( log 2 ( n + 2 ) ) k + 1 / 2 ) \log_2 f_k(n)=\Theta((\log_2(n+2))^{k+1/2}) log 2 f k ( n ) = Θ (( log 2 ( n + 2 ) ) k + 1/2 ) . The same argument with f ( n ) = 2 n f(n)=2^n f ( n ) = 2 n separates every Q P k \mathrm{QP}_k QP k from E X P \mathrm{EXP} EXP .
Corollary 5.8 (A fixed law with class 2 E X P \mathbf{2EXP} 2EXP ) Let J J J have the law (17), Pr ( J = j ) = 1 / ( j ( j + 1 ) ) \Pr(J=j)=1/(j(j+1)) Pr ( J = j ) = 1/ ( j ( j + 1 )) , and let μ \mu μ be the law of K = 2 2 J K=2^{2^J} K = 2 2 J . This law is computable, and P μ , α ∗ = P μ ∗ = 2 E X P \mathrm{P}^{*}_{\mu,\alpha}=\mathrm{P}^{*}_{\mu}=\mathbf{2EXP} P μ , α ∗ = P μ ∗ = 2EXP for every 0 < α < 1 0<\alpha<1 0 < α < 1 .
Proof. Here Pr ( K ≥ T ) = 1 / h ( log 2 T ) \Pr(K\ge T)=1/h(\log_2 T) Pr ( K ≥ T ) = 1/ h ( log 2 T ) for T > 1 T>1 T > 1 , with h h h from Theorem 5.2. The increasing atoms and their rational masses are computable. If q ≥ ℓ α q\ge\ell_\alpha q ≥ ℓ α , q ≥ 1 q\ge1 q ≥ 1 , and m = ⌊ q / ℓ α ⌋ m=\lfloor q/\ell_\alpha\rfloor m = ⌊ q / ℓ α ⌋ , the target 2 2 m 2^{2^m} 2 2 m has tail 1 / m ≥ ℓ α / q 1/m\ge\ell_\alpha/q 1/ m ≥ ℓ α / q , so R α ( q ) ≥ 2 2 ⌊ q / ℓ α ⌋ R_\alpha(q)\ge2^{2^{\lfloor q/\ell_\alpha\rfloor}} R α ( q ) ≥ 2 2 ⌊ q / ℓ α ⌋ . Conversely, for q ≥ 1 q\ge1 q ≥ 1 , an eligible u ≥ 2 u\ge2 u ≥ 2 has log 2 log 2 u ≤ h ( log 2 u ) ≤ q / α \log_2\log_2u\le h(\log_2u)\le q/\alpha log 2 log 2 u ≤ h ( log 2 u ) ≤ q / α , so R α ( q ) ≤ q 2 2 q / α R_\alpha(q)\le q\,2^{2^{q/\alpha}} R α ( q ) ≤ q 2 2 q / α by Theorem 2.10. For a decider within 2 2 p ( n ) 2^{2^{p(n)}} 2 2 p ( n ) transitions, take an allowed integer polynomial p ≥ 1 p\ge1 p ≥ 1 and budget Q ( n ) = ⌈ ℓ α ⌉ p ( n ) Q(n)=\lceil\ell_\alpha\rceil p(n) Q ( n ) = ⌈ ℓ α ⌉ p ( n ) ; its radius reaches that bound. Conversely, simulating a witness through Q ( n ) 2 2 Q ( n ) / α Q(n)2^{2^{Q(n)/\alpha}} Q ( n ) 2 2 Q ( n ) / α steps, with 1 / α 1/\alpha 1/ α replaced by a larger integer and index zero checked separately, gives a doubly exponential-time decider. Polynomial simulation overhead is absorbed in the polynomial exponent. Equality at each confidence also gives the intersection equality.
The effectiveness boundary
Computability of the law bounds its class by decidability. We state the individual-law results here and prove them in Appendix C.
Proposition 5.9 (Decidability for computable laws) If μ \mu μ is computable in the sense of Theorem 2.6, then P μ , α ∗ ⊆ R E C \mathrm{P}^{*}_{\mu,\alpha}\subseteq\mathbf{REC} P μ , α ∗ ⊆ REC for every fixed 0 < α < 1 0<\alpha<1 0 < α < 1 , where R E C \mathbf{REC} REC denotes the decidable languages. No moment, drift, or sampling-cost assumption is used.
Computable masses yield a computable cutoff b t a i l ( q ) b_{\mathrm{tail}}(q) b tail ( q ) for each finite budget q q q . Making Pr ( K > b t a i l ( q ) ) \Pr(K>b_{\mathrm{tail}}(q)) Pr ( K > b tail ( q )) sufficiently small excludes targets beyond q b t a i l ( q ) q b_{\mathrm{tail}}(q) q b tail ( q ) at the given confidence. A decider simulates the witnessing machine through that logical-time bound.
Corollary 5.10 (Computable clocks collectively capture decidability) For every fixed finite nonempty input alphabet and every 0 < α < 1 0<\alpha<1 0 < α < 1 ,
⋃ μ c o m p u t a b l e P μ , α ∗ = R E C , ⋃ μ c o m p u t a b l e P μ ∗ = R E C . \bigcup_{\mu\ \mathrm{computable}}\mathrm{P}^{*}_{\mu,\alpha}=\mathbf{REC},\qquad
\bigcup_{\mu\ \mathrm{computable}}\mathrm{P}^{*}_{\mu}=\mathbf{REC}. μ computable ⋃ P μ , α ∗ = REC , μ computable ⋃ P μ ∗ = REC . The witnessing law may depend on the language.
Proposition 5.11 (An arbitrary law attaining all semidecidable languages) For every fixed finite nonempty input alphabet there is one law μ \mu μ on the positive integers such that, for every 0 < α < 1 0<\alpha<1 0 < α < 1 ,
P μ , α ∗ = P μ ∗ = R E . \mathrm{P}^{*}_{\mu,\alpha}=\mathrm{P}^{*}_{\mu}=\mathbf{RE}. P μ , α ∗ = P μ ∗ = RE . This law is not computable.
The law places mass 1 / ( j ( j + 1 ) ) 1/(j(j+1)) 1/ ( j ( j + 1 )) on a busy-beaver-type halting-time envelope ( E j ) (E_j) ( E j ) ; every recognizer’s accepting indices are then overtaken by a single atom with probability 1 / j 1/j 1/ j . The law is noncomputable by Theorem 5.9, since RE contains undecidable languages.
Backward moves
Deleting backward moves never worsens reachability for iid clocks (Theorem 6.4) or for kernels with monotone forward ability (Theorem 6.3 ). A length-one return can reopen a heavy-sampling phase and change the class from polynomial to exponential time. In the gated family, backward moves are necessary to exceed P (§ 6.3 ).
We use the kernel classes defined after Theorem 2.3, with the clock started at zero. Superscripts on probabilities and expectations specify a kernel parameter when needed; subscripts specify the starting index. For an iid law, we also write C μ , α = C η μ , α = P μ , α ∗ C_{\mu,\alpha}=C_{\eta_\mu,\alpha}=P^*_{\mu,\alpha} C μ , α = C η μ , α = P μ , α ∗ .
Deletion and monotone forward ability
Deleting backward moves preserves the forward-displacement law at each index. A monotone ordering of these laws couples the deleted clock above the original.
Definition 6.1 (Deletion of backward moves) For a stochastic kernel η \eta η on N \mathbb{N} N , define D η D\eta D η by ( D η ) ( i , j ) = η ( i , j ) (D\eta)(i,j)=\eta(i,j) ( D η ) ( i , j ) = η ( i , j ) for j > i j>i j > i , ( D η ) ( i , i ) = ∑ j ≤ i η ( i , j ) (D\eta)(i,i)=\sum_{j\le i}\eta(i,j) ( D η ) ( i , i ) = ∑ j ≤ i η ( i , j ) , and ( D η ) ( i , j ) = 0 (D\eta)(i,j)=0 ( D η ) ( i , j ) = 0 for j < i j<i j < i . We call η \eta η forward-only if D η = η D\eta=\eta D η = η .
Definition 6.2 (Forward ability) For a stochastic kernel η \eta η on N \mathbb{N} N and i ∈ N i\in\mathbb{N} i ∈ N , the forward-displacement law φ i η \varphi_i^\eta φ i η on N \mathbb{N} N is
φ i η ( k ) = η ( i , i + k ) ( k ≥ 1 ) , φ i η ( 0 ) = ∑ j ≤ i η ( i , j ) . \varphi_i^\eta(k)=\eta(i,i+k)\quad(k\ge1),\qquad\varphi_i^\eta(0)=\sum_{j\le i}\eta(i,j). φ i η ( k ) = η ( i , i + k ) ( k ≥ 1 ) , φ i η ( 0 ) = j ≤ i ∑ η ( i , j ) . Thus ( D η ) ( i , i + k ) = φ i η ( k ) (D\eta)(i,i+k)=\varphi_i^\eta(k) ( D η ) ( i , i + k ) = φ i η ( k ) for every k ≥ 0 k\ge0 k ≥ 0 , and η \eta η and D η D\eta D η have the same forward-displacement laws. We say that η \eta η has monotone forward ability if φ i η ⪯ s t φ i ′ η \varphi_i^\eta\preceq_{\mathrm{st}}\varphi_{i'}^\eta φ i η ⪯ st φ i ′ η whenever i ≤ i ′ i\le i' i ≤ i ′ , that is, φ i η ( [ k , ∞ ) ) ≤ φ i ′ η ( [ k , ∞ ) ) \varphi_i^\eta([k,\infty))\le\varphi_{i'}^\eta([k,\infty)) φ i η ([ k , ∞ )) ≤ φ i ′ η ([ k , ∞ )) for all k ∈ N k\in\mathbb{N} k ∈ N .
Monotone forward ability ensures that returns cannot improve reachability.
Theorem 6.3 (Backward moves need non-monotone forward ability) If η \eta η has monotone forward ability , then for all q , T ∈ N q,T\in\mathbb{N} q , T ∈ N , with both clocks started at zero,
Pr η ( τ T ≤ q ) ≤ Pr D η ( τ T ≤ q ) . \Pr_{\eta}(\tau_T\le q)\le\Pr_{D\eta}(\tau_T\le q). η Pr ( τ T ≤ q ) ≤ D η Pr ( τ T ≤ q ) . Consequently C η , α ⊆ C D η , α C_{\eta,\alpha}\subseteq C_{D\eta,\alpha} C η , α ⊆ C D η , α and the crossing budget of η \eta η is at least that of D η D\eta D η , for every 0 < α < 1 0<\alpha<1 0 < α < 1 . If every member of a family F \mathcal{F} F has monotone forward ability, the same inclusion holds for common budgets over F \mathcal{F} F and over { D η : η ∈ F } \{D\eta:\eta\in\mathcal{F}\} { D η : η ∈ F } .
Proof. Write φ i = φ i η \varphi_i=\varphi_i^\eta φ i = φ i η and let F i − 1 ( u ) = min { k : φ i ( [ 0 , k ] ) ≥ u } F_i^{-1}(u)=\min\{k:\varphi_i([0,k])\ge u\} F i − 1 ( u ) = min { k : φ i ([ 0 , k ]) ≥ u } for 0 < u < 1 0<u<1 0 < u < 1 . Monotone forward ability gives F i − 1 ( u ) ≤ F i ′ − 1 ( u ) F_i^{-1}(u)\le F_{i'}^{-1}(u) F i − 1 ( u ) ≤ F i ′ − 1 ( u ) for i ≤ i ′ i\le i' i ≤ i ′ . Let U 1 , U 2 , … U_1,U_2,\ldots U 1 , U 2 , … and V 1 , V 2 , … V_1,V_2,\ldots V 1 , V 2 , … be independent uniform variables on ( 0 , 1 ) (0,1) ( 0 , 1 ) . Put X 0 = Y 0 = 0 X_0=Y_0=0 X 0 = Y 0 = 0 and Y t + 1 = Y t + F Y t − 1 ( U t + 1 ) Y_{t+1}=Y_t+F_{Y_t}^{-1}(U_{t+1}) Y t + 1 = Y t + F Y t − 1 ( U t + 1 ) . Let k = F X t − 1 ( U t + 1 ) k=F_{X_t}^{-1}(U_{t+1}) k = F X t − 1 ( U t + 1 ) . If k ≥ 1 k\ge1 k ≥ 1 put X t + 1 = X t + k X_{t+1}=X_t+k X t + 1 = X t + k . If k = 0 k=0 k = 0 , choose X t + 1 ∈ { 0 , … , X t } X_{t+1}\in\{0,\ldots,X_t\} X t + 1 ∈ { 0 , … , X t } using V t + 1 V_{t+1} V t + 1 , with probabilities η ( X t , j ) / φ X t ( 0 ) \eta(X_t,j)/\varphi_{X_t}(0) η ( X t , j ) / φ X t ( 0 ) ; this case has probability zero when φ X t ( 0 ) = 0 \varphi_{X_t}(0)=0 φ X t ( 0 ) = 0 . Then X X X is a Markov chain with kernel η \eta η and Y Y Y one with kernel D η D\eta D η .
We show X t ≤ Y t X_t\le Y_t X t ≤ Y t by induction. In both cases X t + 1 ≤ X t + F X t − 1 ( U t + 1 ) X_{t+1}\le X_t+F_{X_t}^{-1}(U_{t+1}) X t + 1 ≤ X t + F X t − 1 ( U t + 1 ) , so
X t + 1 ≤ X t + F X t − 1 ( U t + 1 ) ≤ Y t + F Y t − 1 ( U t + 1 ) = Y t + 1 . X_{t+1}\le X_t+F_{X_t}^{-1}(U_{t+1})\le Y_t+F_{Y_t}^{-1}(U_{t+1})=Y_{t+1}. X t + 1 ≤ X t + F X t − 1 ( U t + 1 ) ≤ Y t + F Y t − 1 ( U t + 1 ) = Y t + 1 . The chain Y Y Y is nondecreasing, so max s ≤ t X s ≤ Y t \max_{s\le t}X_s\le Y_t max s ≤ t X s ≤ Y t and { τ T ( X ) ≤ q } ⊆ { τ T ( Y ) ≤ q } \{\tau_T(X)\le q\}\subseteq\{\tau_T(Y)\le q\} { τ T ( X ) ≤ q } ⊆ { τ T ( Y ) ≤ q } . The class and budget statements follow from Theorem 2.4 with the same witnesses, and the common-budget statement follows by taking infima over F \mathcal{F} F . □ \square □
Corollary 6.4 (Independent-proposal deletion principle) For every independent-proposal law μ \mu μ and every 0 < α < 1 0 < \alpha< 1 0 < α < 1 ,
C μ , α ⊆ C μ + , α , \mathcal{C}_{\mu,\alpha} \subseteq\mathcal{C}_{\mu^+,\alpha}, C μ , α ⊆ C μ + , α , where μ + \mu^+ μ + is the law of K + = max ( K , 0 ) K^+ = \max(K,0) K + = max ( K , 0 ) . Consequently, if deleting the negative proposals already yields a class contained in P \mathbf{P} P , adding those iid negative proposals cannot produce a larger class.
Proof. For held kernels, φ η μ \varphi^{\eta_\mu} φ η μ is the law of K + K^+ K + at every position and D η μ = η μ + D\eta_\mu= \eta_{\mu^+} D η μ = η μ + , so Theorem 6.3 applies. Alternatively, use the same proposals for both clocks. The negative-suppressed position is X t + = ∑ s ≤ t K s + X_t^+ = \sum_{s\leq t} K_s^+ X t + = ∑ s ≤ t K s + . Each actual increment of the held walk is at most K s + K_s^+ K s + . By the pathwise bound of Theorem 2.8, X t ≤ X t + X_t \leq X_t^+ X t ≤ X t + for every t t t . Hence τ T ( X + ) ≤ τ T ( X ) \tau_T(X^+) \leq\tau_T(X) τ T ( X + ) ≤ τ T ( X ) pathwise. The same recognizing machine and budget therefore work for the negative-suppressed clock.
Both constructions below violate forward monotonicity: the parity rows alternate a heavy law with bounded steps, while a gate’s doubling exceeds the next row’s bound once g + 1 > w g+1>w g + 1 > w . The next proposition shows that bounded positive displacements cannot raise the class beyond P \mathbf{P} P .
Proposition 6.5 (Bounded forward steps cannot reach beyond P \mathbf{P} P ) For any kernel, position-dependent or not, whose positive displacements are all at most a fixed w w w , C η , α ⊆ P \mathcal{C}_{\eta,\alpha} \subseteq\mathbf{P} C η , α ⊆ P .
Proof. For every path, max t ≤ q X t ≤ w q \max_{t\leq q} X_t \leq wq max t ≤ q X t ≤ w q . A witness can therefore accept a yes-instance only at a logical index at most w Q ( ∣ x ∣ ) wQ(|x|) w Q ( ∣ x ∣ ) . Simulate its deterministic path for this many transitions and reject on timeout. Soundness makes this a polynomial-time decider.
Phase reuse in the parity clock
A return by one restores the phase that samples the heavy tail. Reusing that phase changes the class from polynomial to exponential time.
Let Y = 2 J Y=2^J Y = 2 J , where
Pr ( J = m ) = 1 m ( m + 1 ) , m = 1 , 2 , … (18) \Pr(J=m)=\frac{1}{m(m+1)}, \qquad m=1,2,\ldots \tag*{(18)} Pr ( J = m ) = m ( m + 1 ) 1 , m = 1 , 2 , … (18)
Recall h ( z ) = max { 1 , ⌈ log 2 z ⌉ } h(z)=\max\{1,\lceil\log_2 z\rceil\} h ( z ) = max { 1 , ⌈ log 2 z ⌉} from Theorem 5.2. Telescoping gives, for every real z > 0 z>0 z > 0 ,
Pr ( Y ≥ z ) = 1 h ( z ) . (19) \Pr(Y\geq z)=\frac{1}{h(z)}. \tag*{(19)} Pr ( Y ≥ z ) = h ( z ) 1 . (19)
This is the positive-jump law μ E \mu_E μ E of Theorem 5.2.
Fix β ∈ [ 0 , 1 / 2 ] \beta\in[0,1/2] β ∈ [ 0 , 1/2 ] . Define a time-homogeneous Markov kernel η β \eta_\beta η β on N \mathbb{N} N by the complete transition table in Table 2 .
Table 2. Transition probabilities of the activation kernel η β \eta_\beta η β
In this notation η 0 = D η β \eta_0=D\eta_\beta η 0 = D η β .
The kernel is fixed independently of the input and machine. Its entries are rational for rational β \beta β and computable for computable β \beta β . At positive indices, its rows depend only on parity.
The deletion η 0 \eta_0 η 0 replaces each return by a hold. Theorem 2.7 identifies this parity-dependent kernel as outside the iid held family: its odd and even rows have different probabilities for the same positive displacements. Theorem 6.4 explains why merely adding iid negative proposals cannot produce this increase.
Figure 3. The two phases of the clock η β \eta_\beta η β . From an odd index the clock samples Y = 2 J Y = 2^J Y = 2 J with law μ E \mu_E μ E and moves by 1 + Y 1 + Y 1 + Y ; from a positive even index it moves forward by two or, with probability β \beta β , returns by one. A backward step (dashed) restores the phase in which a fresh heavy jump is available; the index itself keeps growing after each completed cycle
At an odd index a fresh heavy jump lands at an even index; a return by one makes that sample available again. From any state below T T T , two ticks give probability at least β / h ( T ) \beta/h(T) β / h ( T ) of crossing, while the return and the following jump’s + 1 +1 + 1 cancel.
A length-one return changes the parity clock’s class from P to EXP .
Theorem 6.6 (Backward-enabled class transition) For every fixed 0 < α < 1 0 < \alpha< 1 0 < α < 1 and every fixed β ∈ ( 0 , 1 / 2 ] \beta\in(0, 1/2] β ∈ ( 0 , 1/2 ] ,
C η 0 , α = P , C η β , α = E X P . \mathcal{C}_{\eta_0,\alpha} = \mathbf{P}, \qquad\mathcal{C}_{\eta_\beta,\alpha} = \mathbf{EXP}. C η 0 , α = P , C η β , α = EXP . The two clocks differ only in that η β \eta_\beta η β may step back by one from a positive even index, giving a strict increase from P to EXP . At β = 0 \beta= 0 β = 0 only the first odd sampling state is visited, after which the walk stays even. At β > 0 \beta> 0 β > 0 length-one returns repeatedly reopen the odd sampling phase. The increase is phase reuse through accessibility.
Remark 6.7 (Forward re-entry in the parity family) Theorem 6.6 compares η β \eta_\beta η β with its deletion D η β = η 0 D\eta_\beta= \eta_0 D η β = η 0 . It does not show that backward displacement is needed. Replace each move i → i − 1 i \to i - 1 i → i − 1 from a positive even index by i → i + 1 i \to i + 1 i → i + 1 . The resulting kernel is forward-only, and it still enters an odd sampling index with probability β \beta β at each even-state tick. The proofs of Theorems 6.9 and 6.10 apply with only notational changes: from an even index the step to i + 1 i + 1 i + 1 followed by a jump 2 m ≥ T 2^m \ge T 2 m ≥ T crosses T T T , and every positive displacement is still at most 1 + Y t 1 + Y_t 1 + Y t . Its class is therefore EXP as well. Section 6.3 gives a family in which backward displacement is necessary.
The deterministic time hierarchy theorem gives strictness. The proof uses the access rule for a fixed kernel and applies to every real β \beta β in the parameter interval.
Lemma 6.8 For η 0 \eta_0 η 0 , a single copy of Y Y Y can be chosen such that
max t ≤ q X t ≤ 2 q + Y ( q ≥ 0 ) . \max_{t \le q} X_t \le2q + Y \qquad(q \ge0). t ≤ q max X t ≤ 2 q + Y ( q ≥ 0 ) . Moreover E i τ T ≤ 2 T \mathbb{E}_i\tau_T \le2T E i τ T ≤ 2 T for every 0 ≤ i < T 0 \le i < T 0 ≤ i < T .
Proof. Starting at zero, the first two positions are 1 1 1 and 2 + Y 2 + Y 2 + Y ; all later increments are 0 0 0 or 2 2 2 . The maximum bound follows, with an unused independent Y Y Y if q < 2 q < 2 q < 2 .
For the expectation bound, couple each tick with an independent fair Bernoulli variable ξ t \xi_t ξ t . At even indices let ξ t \xi_t ξ t select the + 2 +2 + 2 move. At odd indices and at zero the mandatory positive increment is at least ξ t \xi_t ξ t . The clock’s position is therefore at least ∑ s ≤ t ξ s \sum_{s \le t}\xi_s ∑ s ≤ t ξ s . The time to T T T Bernoulli successes has mean 2 T 2T 2 T and bounds τ T \tau_T τ T . This argument also works from every starting index below T T T . □
Proof of Theorem 6.6 for β = 0 \beta= 0 β = 0 . Fix α \alpha α . Choose a constant b α , t a i l = 2 r b_{\alpha,\mathrm{tail}} = 2^r b α , tail = 2 r with Pr ( Y > b α , t a i l ) = 1 / ( r + 1 ) < α \operatorname{Pr}(Y > b_{\alpha,\mathrm{tail}}) = 1/(r + 1) < \alpha Pr ( Y > b α , tail ) = 1/ ( r + 1 ) < α . If T > 2 q + b α , t a i l T > 2q + b_{\alpha,\mathrm{tail}} T > 2 q + b α , tail , then
Pr ( 0 ) ( τ T ≤ q ) ≤ Pr ( Y > b α , t a i l ) < α . \operatorname{Pr}^{(0)}(\tau_T \le q) \le\operatorname{Pr}(Y > b_{\alpha,\mathrm{tail}}) < \alpha. Pr ( 0 ) ( τ T ≤ q ) ≤ Pr ( Y > b α , tail ) < α .
An accepting index for a witness of budget Q Q Q is consequently at most 2 Q ( ∣ x ∣ ) + b α , t a i l 2Q(|x|)+b_{\alpha,\mathrm{tail}} 2 Q ( ∣ x ∣ ) + b α , tail . Simulating the witnessing machine up to this polynomial bound decides the language. This proves C η 0 , α ⊆ P C_{\eta_0,\alpha} \subseteq\mathbf{P} C η 0 , α ⊆ P .
Conversely, let a polynomial-time decider have logical-time bound H ( n ) H(n) H ( n ) . Theorem 6.8 and Markov’s inequality give Pr ( τ T > q ) ≤ 2 T / q \operatorname{Pr}(\tau_T > q) \le2T/q Pr ( τ T > q ) ≤ 2 T / q . Taking an integer constant k α , l i n ≥ 2 / ( 1 − α ) k_{\alpha,\mathrm{lin}} \ge2/(1-\alpha) k α , lin ≥ 2/ ( 1 − α ) and budget k α , l i n ( H ( n ) + 1 ) k_{\alpha,\mathrm{lin}}(H(n)+1) k α , lin ( H ( n ) + 1 ) supplies completeness; soundness is inherited from the decider. Hence P ⊆ C η 0 , α \mathbf{P} \subseteq C_{\eta_0,\alpha} P ⊆ C η 0 , α .
Lemma 6.9 (Uniform two-tick amplification) Let β > 0 \beta> 0 β > 0 , T ≥ 1 T \ge1 T ≥ 1 , and 0 ≤ i < T 0 \le i < T 0 ≤ i < T . For every integer r ≥ 0 r \ge0 r ≥ 0 ,
Pr i ( τ T > 2 r ) ≤ ( 1 − β h ( T ) ) r ≤ exp ( − β r h ( T ) ) . (20) \operatorname{Pr}_{i}(\tau_T > 2r) \le\left(1-\frac{\beta}{h(T)}\right)^r \le\exp\left(-\frac{\beta r}{h(T)}\right). \tag*{(20)} Pr i ( τ T > 2 r ) ≤ ( 1 − h ( T ) β ) r ≤ exp ( − h ( T ) β r ) . (20) In particular, with q α ( T ) q_\alpha(T) q α ( T ) as in (2) ,
q α ( T ) ≤ 2 ⌈ h ( T ) β ℓ α ⌉ . (21) q_\alpha(T) \le2\left\lceil\frac{h(T)}{\beta}\ell_\alpha\right\rceil. \tag*{(21)} q α ( T ) ≤ 2 ⌈ β h ( T ) ℓ α ⌉ . (21)
Proof. At an odd i < T i < T i < T , a heavy jump with Y ≥ T − i − 1 Y \ge T-i-1 Y ≥ T − i − 1 crosses the target on the next tick; since T − i − 1 ≤ T T-i-1 \le T T − i − 1 ≤ T , this has probability at least 1 / h ( T ) 1/h(T) 1/ h ( T ) (and probability one when T − i − 1 ≤ 0 T-i-1 \le0 T − i − 1 ≤ 0 ). At a positive even i < T i < T i < T , the two-event sequence “backward by one, then a heavy jump” has probability β / h ( T ) \beta/h(T) β / h ( T ) and crosses within two ticks: from i − 1 i-1 i − 1 (odd) the jump 2 m ≥ T 2^m \ge T 2 m ≥ T reaches at least i + 2 m ≥ T i+2^m \ge T i + 2 m ≥ T . At zero the first tick goes to 1 and the second offers a heavy jump, again with probability 1 / h ( T ) 1/h(T) 1/ h ( T ) . Thus from every state below T T T the next two ticks cross with probability at least β / h ( T ) \beta/h(T) β / h ( T ) .
Condition successively on the past at the deterministic boundaries 0 , 2 , … , 2 r 0,2,\ldots,2r 0 , 2 , … , 2 r . On failure the position is still below T T T , so the same conditional bound applies. Iteration gives the first inequality; 1 − u ≤ e − u 1-u \le e^{-u} 1 − u ≤ e − u gives the second. Solving for failure probability 1 − α 1-\alpha 1 − α yields (20).
For a decider with logical bound H ( n ) = 2 p ( n ) H(n)=2^{p(n)} H ( n ) = 2 p ( n ) , take an integer polynomial p ≥ 1 p \ge1 p ≥ 1 and an integer k α , β ≥ max { 1 , β − 1 ℓ α } k_{\alpha,\beta} \ge\max\{1,\beta^{-1}\ell_\alpha\} k α , β ≥ max { 1 , β − 1 ℓ α } . The allowed budget Q ( n ) = 2 k α , β p ( n ) Q(n)=2k_{\alpha,\beta}p(n) Q ( n ) = 2 k α , β p ( n ) contains r = k α , β p ( n ) r=k_{\alpha,\beta}p(n) r = k α , β p ( n ) two-tick blocks. Since h ( H ( n ) ) = p ( n ) h(H(n))=p(n) h ( H ( n )) = p ( n ) , (19) bounds failure by e − β k α , β ≤ 1 − α e^{-\beta k_{\alpha,\beta}} \le1-\alpha e − β k α , β ≤ 1 − α . The same machine is sound, so E X P ⊆ C η β , α \mathbf{EXP} \subseteq C_{\eta_\beta,\alpha} EXP ⊆ C η β , α .
Lemma 6.10 (Exponential ceiling) For every β ∈ [ 0 , 1 / 2 ] \beta\in[0,1/2] β ∈ [ 0 , 1/2 ] , q ≥ 1 q \ge1 q ≥ 1 , and T > q T > q T > q ,
Pr ( β ) ( τ T ≤ q ) ≤ q h ( T − q q ) . (22) \operatorname{Pr}^{(\beta)}(\tau_T \le q) \le\frac{q}{h\left(\frac{T-q}{q}\right)}. \tag*{(22)} Pr ( β ) ( τ T ≤ q ) ≤ h ( q T − q ) q . (22) Thus success probability at least α \alpha α implies
T ≤ q ( 1 + 2 q / α ) . (23) T \le q\left(1+2^{q/\alpha}\right). \tag*{(23)} T ≤ q ( 1 + 2 q / α ) . (23)
Proof. Choose iid copies Y 1 , … , Y q Y_1,\ldots,Y_q Y 1 , … , Y q , one for each tick, using a copy for the actual heavy jump only when the current index is odd. Every positive displacement is at most 1 + Y t 1+Y_t 1 + Y t : this covers the initial + 1 +1 + 1 , the even-state + 2 +2 + 2 , and the odd-state 1 + 2 m 1+2^m 1 + 2 m . Consequently
max t ≤ q X t ≤ q + ∑ t = 1 q Y t . \max_{t\le q} X_t \le q+\sum_{t=1}^{q}Y_t. t ≤ q max X t ≤ q + t = 1 ∑ q Y t .
If the target is reached, some Y t ≥ ( T − q ) / q Y_t \ge(T-q)/q Y t ≥ ( T − q ) / q ; otherwise the sum is less than T − q T-q T − q . The union bound and (18) prove (21). Success at level α \alpha α forces h ( ( T − q ) / q ) ≤ q / α h((T-q)/q) \le q/\alpha h (( T − q ) / q ) ≤ q / α , hence ( T − q ) / q ≤ 2 q / α (T-q)/q \le2^{q/\alpha} ( T − q ) / q ≤ 2 q / α . If T ≤ q T \le q T ≤ q , (22) also holds.
Completion of Theorem 6.6. For a witness of polynomial budget Q Q Q , every accepting index is bounded by Q ( n ) ( 1 + 2 Q ( n ) / α ) Q(n)(1+2Q(n)/\alpha) Q ( n ) ( 1 + 2 Q ( n ) / α ) when Q ( n ) ≥ 1 Q(n)\ge1 Q ( n ) ≥ 1 . Replace 1 / α 1/\alpha 1/ α by any larger hardcoded integer and simulate the deterministic path through that bound, rejecting on timeout. This is an exponential-time decider, by soundness. Thus C η β , α ⊆ E X P C_{\eta_\beta,\alpha} \subseteq\mathrm{EXP} C η β , α ⊆ EXP ; the preceding amplification argument proves the opposite inclusion for β > 0 \beta>0 β > 0 .
Recall q α ( T ) q_\alpha(T) q α ( T ) from (2); we write q α ( β ) q_\alpha^{(\beta)} q α ( β ) for the clock η β \eta_\beta η β .
The same phase reuse gives exact leading constants for the crossing budget.
Theorem 6.11 (Sharp linear and logarithmic scales) Fix 0 < α < 1 0<\alpha<1 0 < α < 1 . As integer T → ∞ T\to\infty T → ∞ ,
q α ( 0 ) ( T ) T ⟶ 1 , \frac{q_\alpha^{(0)}(T)}{T} \longrightarrow1, T q α ( 0 ) ( T ) ⟶ 1 , q α ( β ) ( T ) log 2 T ⟶ ( 1 + 1 β ) ℓ α ( β > 0 ) . \frac{q_\alpha^{(\beta)}(T)}{\log_2 T} \longrightarrow\left(1+\frac{1}{\beta}\right)\ell_\alpha\qquad(\beta>0). log 2 T q α ( β ) ( T ) ⟶ ( 1 + β 1 ) ℓ α ( β > 0 ) .
Proof. For β = 0 \beta=0 β = 0 and q ≥ 2 q\ge2 q ≥ 2 , the clock is nondecreasing and
X q = 2 + Y + 2 F q − 2 , F q − 2 ∼ Bin ( q − 2 , 1 / 2 ) . X_q = 2+Y+2F_{q-2}, \qquad F_{q-2}\sim\operatorname{Bin}(q-2,1/2). X q = 2 + Y + 2 F q − 2 , F q − 2 ∼ Bin ( q − 2 , 1/2 ) .
For q = ⌊ c T ⌋ q=\lfloor cT\rfloor q = ⌊ c T ⌋ , Y / T → 0 Y/T\to0 Y / T → 0 in probability and 2 F q − 2 / T → c 2F_{q-2}/T\to c 2 F q − 2 / T → c in probability. Thus the crossing probability tends to zero if c < 1 c<1 c < 1 , and to one if c > 1 c>1 c > 1 . Squeezing the fixed-confidence quantile between these budgets gives the first limit.
For β > 0 \beta>0 β > 0 , the first heavy jump completes at tick 2. After the j j j th heavy jump, the number G j G_j G j of even-state ticks up to and including the next backward move has distribution Geom ( β ) \operatorname{Geom}(\beta) Geom ( β ) on { 1 , 2 , … } \{1,2,\ldots\} { 1 , 2 , … } . One further tick completes the next heavy jump. The cycle lengths ϱ j = G j + 1 \varrho_j=G_j+1 ϱ j = G j + 1 are iid, with mean d β = 1 + β − 1 d_\beta=1+\beta^{-1} d β = 1 + β − 1 . They can be generated independently of the samples Y 1 , Y 2 , … Y_1,Y_2,\ldots Y 1 , Y 2 , … : every even-state row has the same probabilities, and each heavy displacement changes an odd index to an even one regardless of its magnitude.
Let N q N_q N q count heavy samples completed by tick q q q . It is zero for q < 2 q<2 q < 2 , and for q ≥ 2 q\ge2 q ≥ 2 ,
N q = 1 + max { m ≥ 0 : ∑ j = 1 m ϱ j ≤ q − 2 } . N_q = 1+\max\left\{m\ge0:\sum_{j=1}^{m}\varrho_j\le q-2\right\}. N q = 1 + max { m ≥ 0 : j = 1 ∑ m ϱ j ≤ q − 2 } .
With S m ϱ = ∑ j = 1 m ϱ j S_m^\varrho=\sum_{j=1}^{m}\varrho_j S m ϱ = ∑ j = 1 m ϱ j and m = N q − 1 m=N_q-1 m = N q − 1 , one has S m ϱ ≤ q − 2 < S m + 1 ϱ S_m^\varrho\le q-2<S_{m+1}^\varrho S m ϱ ≤ q − 2 < S m + 1 ϱ . The strong law for the integrable cycle lengths [18 Theorem 2.4.1] implies S m ϱ / m → d β S_m^\varrho/m\to d_\beta S m ϱ / m → d β and hence N q / q → 1 / d β N_q/q\to1/d_\beta N q / q → 1/ d β almost surely. This count is independent of all heavy samples.
At each completed-cycle endpoint the return − 1 -1 − 1 and the next forward + 1 +1 + 1 cancel. The position is 2 + ∑ r ≤ N q Y r 2+\sum_{r\le N_q}Y_r 2 + ∑ r ≤ N q Y r plus the completed even-state + 2 +2 + 2 steps. At a pending odd state it is one less than the preceding even position. A completed sample Y r ≥ T Y_r\ge T Y r ≥ T therefore forces a hit, while the running maximum is at most 2 q + ∑ r ≤ N q Y r 2q+\sum_{r\le N_q}Y_r 2 q + ∑ r ≤ N q Y r . Since N q ≤ q N_q\le q N q ≤ q , a hit at T > 2 q T>2q T > 2 q forces some completed sample to be at least ( T − 2 q ) / q (T-2q)/q ( T − 2 q ) / q . Independence of N q N_q N q and the samples gives, for T > 2 q T>2q T > 2 q and q ≥ 1 q\ge1 q ≥ 1 ,
1 − E ( 1 − 1 h ( T ) ) N q ≤ Pr ( τ T ≤ q ) ≤ 1 − E ( 1 − 1 h ( ( T − 2 q ) / q ) ) N q . (24) 1-\mathbb{E}\left(1-\frac{1}{h(T)}\right)^{N_q}
\le\Pr(\tau_T\le q)
\le
1-\mathbb{E}\left(1-\frac{1}{h((T-2q)/q)}\right)^{N_q}. \tag*{(24)} 1 − E ( 1 − h ( T ) 1 ) N q ≤ Pr ( τ T ≤ q ) ≤ 1 − E ( 1 − h (( T − 2 q ) / q ) 1 ) N q . (24)
When N q = 0 N_q=0 N q = 0 the conditional failure probability is one, including the convention 0 0 = 1 0^0=1 0 0 = 1 .
Put L T = log 2 T L_T=\log_2 T L T = log 2 T and q = ⌊ c L T ⌋ q=\lfloor cL_T\rfloor q = ⌊ c L T ⌋ , c > 0 c>0 c > 0 fixed. For large T T T the sandwich applies, and both denominators H T = h ( T ) H_T=h(T) H T = h ( T ) and H T = h ( ( T − 2 q ) / q ) H_T=h((T-2q)/q) H T = h (( T − 2 q ) / q ) satisfy H T / L T → 1 H_T/L_T\to1 H T / L T → 1 : subtracting 2 q 2q 2 q and dividing by q q q changes the logarithm by only O ( log L T ) O(\log L_T) O ( log L T ) . For either denominator, N q / L T → c / d β N_q/L_T\to c/d_\beta N q / L T → c / d β in probability and N q / H T 2 ≤ q / H T 2 → 0 N_q/H_T^2\le q/H_T^2\to0 N q / H T 2 ≤ q / H T 2 → 0 . Since ln \ln ln is natural,
N q ln ( 1 − 1 / H T ) = − N q / H T + O ( N q / H T 2 ) ⟶ − c / d β N_q\ln(1-1/H_T)=-N_q/H_T+O(N_q/H_T^2)\longrightarrow-c/d_\beta N q ln ( 1 − 1/ H T ) = − N q / H T + O ( N q / H T 2 ) ⟶ − c / d β
in probability. The powers lie in [ 0 , 1 ] [0,1] [ 0 , 1 ] . For any error margin δ > 0 \delta> 0 δ > 0 , their expected distance from e − c / d β e^{-c/d^{\beta}} e − c / d β is at most δ + Pr ( ∣ ( 1 − 1 / H T ) N q − e − c / d β ∣ > δ ) \delta+ \Pr\left(\left|(1 - 1/H_T)^{N_q} - e^{-c/d^{\beta}}\right| > \delta\right) δ + Pr ( ( 1 − 1/ H T ) N q − e − c / d β > δ ) . It tends to at most δ \delta δ , and then to zero as δ ↓ 0 \delta\downarrow0 δ ↓ 0 . Thus both bounds converge and
Pr ( τ T ≤ ⌊ c log 2 T ⌋ ) ⟶ 1 − e − c / d β . \Pr(\tau_T \le\lfloor c\log_2 T\rfloor) \longrightarrow1 - e^{-c/d^{\beta}}. Pr ( τ T ≤ ⌊ c log 2 T ⌋) ⟶ 1 − e − c / d β .
The target logarithm is base two, while ℓ α = − ln ( 1 − α ) \ell_\alpha= -\ln(1-\alpha) ℓ α = − ln ( 1 − α ) uses the natural logarithm. Set c ∗ = d β ℓ α c_* = d^{\beta}\ell_\alpha c ∗ = d β ℓ α and fix 0 < ε < c ∗ 0 < \varepsilon< c_* 0 < ε < c ∗ . The limiting probabilities at c ∗ − ε c_*-\varepsilon c ∗ − ε and c ∗ + ε c_*+\varepsilon c ∗ + ε have strictly positive gaps below and above α \alpha α . Choose probability-error margins smaller than half those gaps before taking T T T large. The actual probabilities are then respectively below and above α \alpha α . Monotonicity in the integer budget places q α ( β ) ( T ) q_{\alpha}^{(\beta)}(T) q α ( β ) ( T ) between the corresponding floor budgets; dividing by L T L_T L T removes their at-most-one rounding errors. Letting ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 proves the second limit.
Backward steps are necessary in a gated family
Sparse gates provide the only large forward moves. We now give a family of position-dependent clocks in which no forward-only member exceeds P \mathbf{P} P , however its bounded forward moves are arranged, while steps back by one reach E X P \mathbf{EXP} EXP .
Definition 6.12 (Gated family) Let Λ = { 2 k : k ≥ 1 } \Lambda= \{2^k : k \ge1\} Λ = { 2 k : k ≥ 1 } be the set of gates . For an integer w ≥ 1 w \ge1 w ≥ 1 , let K w \mathcal{K}_w K w be the set of time-homogeneous stochastic kernels η \eta η on N \mathbb{N} N such that
(a) η ( g , 2 g + 1 ) = 1 \eta(g,2g+1)=1 η ( g , 2 g + 1 ) = 1 for every gate g ∈ Λ g \in\Lambda g ∈ Λ ;
(b) for every i ∉ Λ i \notin\Lambda i ∈ / Λ , η ( i , j ) = 0 \eta(i,j)=0 η ( i , j ) = 0 whenever j > i + w j>i+w j > i + w .
At a non-gate index the row may depend on i i i arbitrarily and may contain backward moves of any length; only its forward displacement is bounded. A gate doubles the index and lands one step beyond the next gate 2 g 2g 2 g . The forward-only members of K w \mathcal{K}_w K w include every kernel obtained from a member by deleting its backward moves, or by replacing them with forward moves of length at most w w w .
For β ∈ ( 0 , 1 ] \beta\in(0,1] β ∈ ( 0 , 1 ] , define η β g a t e ∈ K 1 \eta^{\mathrm{gate}}_\beta\in\mathcal{K}_1 η β gate ∈ K 1 as follows. Put η β g a t e ( 0 , 1 ) = η β g a t e ( 1 , 2 ) = 1 \eta^{\mathrm{gate}}_\beta(0,1)=\eta^{\mathrm{gate}}_\beta(1,2)=1 η β gate ( 0 , 1 ) = η β gate ( 1 , 2 ) = 1 and η β g a t e ( g , 2 g + 1 ) = 1 \eta^{\mathrm{gate}}_\beta(g,2g+1)=1 η β gate ( g , 2 g + 1 ) = 1 for g ∈ Λ g \in\Lambda g ∈ Λ . At each landing index s m = 2 m + 1 s_m=2^m+1 s m = 2 m + 1 with m ≥ 2 m \ge2 m ≥ 2 ,
η β g a t e ( s m , s m − 1 ) = β , η β g a t e ( s m , s m ) = 1 − β , \eta^{\mathrm{gate}}_\beta(s_m,s_m-1)=\beta,\qquad\eta^{\mathrm{gate}}_\beta(s_m,s_m)=1-\beta, η β gate ( s m , s m − 1 ) = β , η β gate ( s m , s m ) = 1 − β ,
and η β g a t e ( i , i + 1 ) = 1 \eta^{\mathrm{gate}}_\beta(i,i+1)=1 η β gate ( i , i + 1 ) = 1 at every other index. Its only backward moves are steps back by one, from s m s_m s m to the gate 2 m 2^m 2 m .
In this family, every forward-only clock stays polynomial, and a step back accesses exponential time.
Theorem 6.13 (Backward steps are necessary in the gated family) Fix an integer w ≥ 1 w \ge1 w ≥ 1 and 0 < α < 1 0 < \alpha< 1 0 < α < 1 .
(i) If η ∈ K w \eta\in\mathcal{K}_w η ∈ K w is forward-only , then surely max t ≤ q X t ≤ 3 ( w + 1 ) q \max_{t \le q} X_t \le3(w+1)q max t ≤ q X t ≤ 3 ( w + 1 ) q for every q ∈ N q \in\mathbb{N} q ∈ N . Hence q α ( T ) ≥ T / ( 3 ( w + 1 ) ) q_\alpha(T) \ge T/(3(w+1)) q α ( T ) ≥ T / ( 3 ( w + 1 )) for every T ≥ 1 T \ge1 T ≥ 1 , and C η , α ⊆ P C_{\eta,\alpha} \subseteq\mathbf{P} C η , α ⊆ P . Equality holds for the member of K 1 \mathcal{K}_1 K 1 that moves by + 1 +1 + 1 at every non-gate index.
(ii) Every η ∈ K w \eta\in\mathcal{K}_w η ∈ K w satisfies X t ≤ w ( 2 t − 1 ) X_t \le w(2^t-1) X t ≤ w ( 2 t − 1 ) surely , so C η , α ⊆ E X P C_{\eta,\alpha} \subseteq\mathbf{EXP} C η , α ⊆ EXP .
(iii) For every β ∈ ( 0 , 1 ] \beta\in(0,1] β ∈ ( 0 , 1 ] , C η β g a t e , α = E X P C_{\eta^{\mathrm{gate}}_\beta,\alpha}=\mathbf{EXP} C η β gate , α = EXP , and the crossing budget of η β g a t e \eta^{\mathrm{gate}}_\beta η β gate satisfies
q α ( T ) log 2 T ⟶ 1 + 1 β ( T → ∞ ) . \frac{q_\alpha(T)}{\log_2 T}\longrightarrow1+\frac{1}{\beta}\qquad(T\to\infty). log 2 T q α ( T ) ⟶ 1 + β 1 ( T → ∞ ) .
In particular, D η β g a t e \mathrm{D}\eta_{\beta}^{\mathrm{gate}} D η β gate and the kernel obtained from η β g a t e \eta_{\beta}^{\mathrm{gate}} η β gate by replacing each step s m → s m − 1 s_m \to s_m - 1 s m → s m − 1 with s m → s m + 1 s_m \to s_m + 1 s m → s m + 1 are forward-only members of K 1 \mathcal{K}_1 K 1 , so their classes are contained in P \mathbf{P} P . Within K w \mathcal{K}_w K w , a class not contained in P \mathbf{P} P requires backward moves, and steps back by one suffice for E X P \mathbf{EXP} EXP .
Proof. (i) A forward-only path is nondecreasing. A gate g g g moves to 2 g + 1 > g 2g+1>g 2 g + 1 > g at the next tick, so each gate is occupied at no more than one tick. Fix t t t and list the gates occupied before tick t t t as g 1 < ⋯ < g N g_1<\cdots<g_N g 1 < ⋯ < g N ; thus N ≤ t N\le t N ≤ t . Put λ 0 = 0 \lambda_0=0 λ 0 = 0 and λ r = 2 g r + 1 \lambda_r=2g_r+1 λ r = 2 g r + 1 . Between the departure from g r − 1 g_{r-1} g r − 1 (or tick zero when r = 1 r=1 r = 1 ) and the arrival at g r g_r g r , every move starts at a non-gate index, so its displacement lies in [ 0 , w ] [0,w] [ 0 , w ] . Let ω r = g r − λ r − 1 \omega_r=g_r-\lambda_{r-1} ω r = g r − λ r − 1 be their total. For r = 1 r=1 r = 1 , ω 1 = g 1 ≥ g 1 / 2 − 1 \omega_1=g_1\ge g_1/2-1 ω 1 = g 1 ≥ g 1 /2 − 1 . For r ≥ 2 r\ge2 r ≥ 2 , write g r − 1 = 2 j g_{r-1}=2^j g r − 1 = 2 j and g r = 2 k g_r=2^k g r = 2 k . Since g r > λ r − 1 = 2 j + 1 + 1 g_r>\lambda_{r-1}=2^{j+1}+1 g r > λ r − 1 = 2 j + 1 + 1 , we have k ≥ j + 2 k\ge j+2 k ≥ j + 2 , so λ r − 1 ≤ 2 k − 1 + 1 \lambda_{r-1}\le2^{k-1}+1 λ r − 1 ≤ 2 k − 1 + 1 and ω r ≥ g r / 2 − 1 \omega_r\ge g_r/2-1 ω r ≥ g r /2 − 1 . In both cases the gain at the r r r th gate satisfies
( 2 g r + 1 ) − g r = g r + 1 ≤ 2 ω r + 3. (2g_r+1)-g_r=g_r+1\le2\omega_r+3. ( 2 g r + 1 ) − g r = g r + 1 ≤ 2 ω r + 3.
Let Ω t \Omega_t Ω t be the total displacement of the non-gate moves made before tick t t t . Then Ω t ≤ w t \Omega_t\le wt Ω t ≤ w t , and the stretches counted by the ω r \omega_r ω r are disjoint, so ∑ r = 1 N ω r ≤ Ω t \sum_{r=1}^{N}\omega_r\le\Omega_t ∑ r = 1 N ω r ≤ Ω t . Therefore
X t = Ω t + ∑ r = 1 N ( g r + 1 ) ≤ 3 Ω t + 3 N ≤ 3 ( w + 1 ) t . X_t=\Omega_t+\sum_{r=1}^{N}(g_r+1)\le3\Omega_t+3N\le3(w+1)t. X t = Ω t + r = 1 ∑ N ( g r + 1 ) ≤ 3 Ω t + 3 N ≤ 3 ( w + 1 ) t .
Since the path is nondecreasing, this bounds max t ≤ q X t \max_{t\le q}X_t max t ≤ q X t . Thus Pr ( τ T ≤ q ) = 0 \Pr(\tau_T\le q)=0 Pr ( τ T ≤ q ) = 0 whenever T > 3 ( w + 1 ) q T>3(w+1)q T > 3 ( w + 1 ) q , which gives the budget bound. A witness with budget Q Q Q accepts every yes-instance at an index at most 3 ( w + 1 ) Q ( ∣ x ∣ ) 3(w+1)Q(|x|) 3 ( w + 1 ) Q ( ∣ x ∣ ) . Simulating through that index, as in Theorem 6.5, decides the language in polynomial time. For the kernel that moves by + 1 +1 + 1 at every non-gate index, every move increases the index, so X t ≥ t X_t\ge t X t ≥ t surely. A decider with polynomial logical bound H H H is therefore observed by tick H ( n ) H(n) H ( n ) with probability one, and P ⊆ C η , α \mathbf{P}\subseteq\mathcal{C}_{\eta,\alpha} P ⊆ C η , α .
(ii) At a gate X t + 1 = 2 X t + 1 X_{t+1}=2X_t+1 X t + 1 = 2 X t + 1 , and elsewhere X t + 1 ≤ X t + w X_{t+1}\le X_t+w X t + 1 ≤ X t + w . Both give X t + 1 ≤ 2 X t + w X_{t+1}\le2X_t+w X t + 1 ≤ 2 X t + w , hence X t ≤ w ( 2 t − 1 ) X_t\le w(2^t-1) X t ≤ w ( 2 t − 1 ) . A witness with budget Q Q Q accepts at indices at most w 2 Q ( ∣ x ∣ ) w2^{Q(|x|)} w 2 Q ( ∣ x ∣ ) , and simulating through that bound gives an exponential-time decider.
(iii) Write η = η β g a t e \eta=\eta_{\beta}^{\mathrm{gate}} η = η β gate and d β = 1 + β − 1 d_\beta=1+\beta^{-1} d β = 1 + β − 1 . From zero the path visits 0 , 1 , 2 , s 2 = 5 0,1,2,s_2=5 0 , 1 , 2 , s 2 = 5 at ticks 0 , 1 , 2 , 3 0,1,2,3 0 , 1 , 2 , 3 . For r ≥ 1 r\ge1 r ≥ 1 , the clock holds at s r + 1 s_{r+1} s r + 1 until it steps to the gate 2 r + 1 2^{r+1} 2 r + 1 , which happens after G r G_r G r ticks, and the next tick moves it to s r + 2 s_{r+2} s r + 2 . Every landing row has the same probabilities, so G 1 , G 2 , … G_1,G_2,\ldots G 1 , G 2 , … are iid with law Geom ( β ) \operatorname{Geom}(\beta) Geom ( β ) on { 1 , 2 , … } \{1,2,\ldots\} { 1 , 2 , … } . The arrival tick at s m s_m s m is therefore
σ m = 3 + ∑ r = 1 m − 2 ( G r + 1 ) ( m ≥ 2 ) . \sigma_m=3+\sum_{r=1}^{m-2}(G_r+1)\qquad(m\ge2). σ m = 3 + r = 1 ∑ m − 2 ( G r + 1 ) ( m ≥ 2 ) .
For m ≥ 3 m\ge3 m ≥ 3 , before σ m \sigma_m σ m every visited index is at most s m − 1 s_{m-1} s m − 1 . For an integer T ≥ 6 T\ge6 T ≥ 6 , put m ( T ) = ⌈ log 2 ( T − 1 ) ⌉ ≥ 3 m(T)=\lceil\log_2(T-1)\rceil\ge3 m ( T ) = ⌈ log 2 ( T − 1 )⌉ ≥ 3 . Then s m ( T ) − 1 < T ≤ s m ( T ) s_{m(T)-1}<T\le s_{m(T)} s m ( T ) − 1 < T ≤ s m ( T ) , so τ T = σ m ( T ) \tau_T=\sigma_{m(T)} τ T = σ m ( T ) . By the strong law of large numbers [18 Theorem 2.4.1] , σ m / m → d β \sigma_m/m\to d_\beta σ m / m → d β almost surely, and m ( T ) / log 2 T → 1 m(T)/\log_2T\to1 m ( T ) / log 2 T → 1 . Hence τ T / log 2 T → d β \tau_T/\log_2T\to d_\beta τ T / log 2 T → d β almost surely, and so in probability. For 0 < ε < d β 0<\varepsilon<d_\beta 0 < ε < d β , the crossing probabilities at the budgets [ ( d β − ε ) log 2 T ] [(d_\beta-\varepsilon)\log_2T] [( d β − ε ) log 2 T ] and [ ( d β + ε ) log 2 T ] [(d_\beta+\varepsilon)\log_2T] [( d β + ε ) log 2 T ] tend to 0 0 0 and 1 1 1 , so for large T T T the quantile q α ( T ) q_\alpha(T) q α ( T ) lies between these budgets. Dividing by log 2 T \log_2T log 2 T and letting ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 proves the limit; it does not depend on α \alpha α .
For E X P ⊆ C η , α \mathbf{EXP}\subseteq\mathcal{C}_{\eta,\alpha} EXP ⊆ C η , α , let a decider run within H ( n ) = 2 p ( n ) H(n)=2^{p(n)} H ( n ) = 2 p ( n ) transitions, where p p p is an integer polynomial with p ≥ 2 p\ge2 p ≥ 2 (after enlarging the bound). Since s p ( n ) ≥ 2 p ( n ) s_{p(n)}\ge2^{p(n)} s p ( n ) ≥ 2 p ( n ) , we have τ H ( n ) ≤ σ p ( n ) \tau_{H(n)}\le\sigma_{p(n)} τ H ( n ) ≤ σ p ( n ) , and E σ p ( n ) ≤ 3 + d β p ( n ) ≤ ( 3 + d β ) p ( n ) \mathbb{E}\sigma_{p(n)}\le3+d_\beta p(n)\le(3+d_\beta)p(n) E σ p ( n ) ≤ 3 + d β p ( n ) ≤ ( 3 + d β ) p ( n ) . Choose an integer k α , β ≥ ( 3 + d β ) / ( 1 − α ) k_{\alpha,\beta}\ge(3+d_\beta)/(1-\alpha) k α , β ≥ ( 3 + d β ) / ( 1 − α ) and put Q ( n ) = k α , β p ( n ) Q(n)=k_{\alpha,\beta}p(n) Q ( n ) = k α , β p ( n ) . Markov's inequality gives Pr ( τ H ( n ) > Q ( n ) ) ≤ 1 − α \Pr(\tau_{H(n)}>Q(n))\le1-\alpha Pr ( τ H ( n ) > Q ( n )) ≤ 1 − α . Accepting indices are at most H ( n ) H(n) H ( n ) , crossing times are monotone in the target, and the decider is sound, so ( M , Q ) (M,Q) ( M , Q ) is a witness. Part (ii) gives the reverse inclusion.
The final assertions follow from part (i), since both modified kernels are forward-only members of K 1 \mathcal{K}_1 K 1 .
Remark 6.14 (Three levels of backward power) For iid clocks, deletion never lowers the class (Theorem 6.4). With fixed parity rows, deletion lowers EXP to P , although forward re-entry also restores EXP (Theorems 6.6 and 6.7). In the gated family every forward-only clock stays in P , so exceeding it requires backward steps (Theorem 6.13). A forward clock overshoots each gate and pays to walk to the next; a step back reopens the gate just passed. These clocks reuse indices of one path, without causal consistency or a physical speedup. For example, at β = α = 1 / 2 \beta=\alpha=1/2 β = α = 1/2 and T = 2 n 2 T=2^{n^2} T = 2 n 2 , Theorem 6.11 gives budgets asymptotic to 3 ( ln 2 ) n 2 3(\ln2)n^2 3 ( ln 2 ) n 2 with returns and 2 n 2 2^{n^2} 2 n 2 without them. Such power requires failure of monotone forward ability (Theorem 6.3 ) and persists with sparse, rare returns (§ 6.4 ).
Sparse and rare backward moves
Returns can be supported on a density-zero set of indices. This is sparsity in index space, not a statement that backward moves have zero frequency along a trajectory. In the doubling family the return probabilities may vanish at the specified landing indices. Squaring gates make the support sparser while giving doubly exponential reach.
For a kernel η \eta η on N \mathbb{N} N , let
Back ( η ) = { i ∈ N : η ( i , j ) > 0 for some j < i } \operatorname{Back}(\eta)=\{i\in\mathbb{N}:\eta(i,j)>0\text{ for some }j<i\} Back ( η ) = { i ∈ N : η ( i , j ) > 0 for some j < i }
be its set of backward indices . A set B ⊆ N B\subseteq\mathbb{N} B ⊆ N has natural density zero if ∣ B ∩ [ 0 , N ) ∣ / N → 0 \lvert B\cap[0,N)\rvert/N\to0 ∣ B ∩ [ 0 , N )∣ / N → 0 . A backward move of probability zero occurs at no tick almost surely, and the empty set is the only counting-measure-null subset of N \mathbb{N} N . The relevant notions of rarity are therefore density zero of Back ( η ) \operatorname{Back}(\eta) Back ( η ) in index space and small backward probabilities at its indices. We use Back ( η ) \operatorname{Back}(\eta) Back ( η ) to distinguish the index set from the parameter set B \mathcal{B} B in Appendix § 14 .
Corollary 6.15 (Density-zero backward indices suffice) For every β ∈ ( 0 , 1 ] \beta\in(0,1] β ∈ ( 0 , 1 ] , Back ( η β g a t e ) = { 2 m + 1 : m ≥ 2 } \operatorname{Back}(\eta^{\mathrm{gate}}_{\beta})=\{2^m+1:m\ge2\} Back ( η β gate ) = { 2 m + 1 : m ≥ 2 } , so ∣ Back ( η β g a t e ) ∩ [ 0 , N ) ∣ ≤ log 2 N \lvert\operatorname{Back}(\eta^{\mathrm{gate}}_{\beta})\cap[0,N)\rvert\le\log_2N ∣ Back ( η β gate ) ∩ [ 0 , N )∣ ≤ log 2 N for N ≥ 2 N\ge2 N ≥ 2 , and this set has natural density zero. Nevertheless, C β , α g a t e = E X P C^{\mathrm{gate}}_{\beta,\alpha}=\mathbf{EXP} C β , α gate = EXP for every 0 < α < 1 0<\alpha<1 0 < α < 1 , whereas every η ∈ K w \eta\in\mathcal{K}_w η ∈ K w with Back ( η ) = ∅ \operatorname{Back}(\eta)=\varnothing Back ( η ) = ∅ has C η , α ⊆ P C_{\eta,\alpha}\subseteq\mathbf{P} C η , α ⊆ P .
Proof. The rows of η β g a t e \eta^{\mathrm{gate}}_{\beta} η β gate with mass below the diagonal are exactly those at s m = 2 m + 1 s_m=2^m+1 s m = 2 m + 1 , m ≥ 2 m\ge2 m ≥ 2 . There are at most log 2 N \log_2N log 2 N such indices below N N N . The class statements are Theorem 6.13(i) and (iii).
Proposition 6.16 (Vanishing backward probability) Fix c ≥ 0 c\ge0 c ≥ 0 . Define η ( c ) g a t e ∈ K 1 \eta^{\mathrm{gate}}_{(c)}\in\mathcal{K}_1 η ( c ) gate ∈ K 1 like η β g a t e \eta^{\mathrm{gate}}_{\beta} η β gate except that at the landing index s m = 2 m + 1 s_m=2^m+1 s m = 2 m + 1 (m ≥ 2 m\ge2 m ≥ 2 ) it moves to s m − 1 s_m-1 s m − 1 with probability β m = m − c \beta_m=m^{-c} β m = m − c and holds with probability 1 − β m 1-\beta_m 1 − β m . Thus the backward probability at s m s_m s m equals ( log 2 ( s m − 1 ) ) − c (\log_2(s_m-1))^{-c} ( log 2 ( s m − 1 ) ) − c , which tends to zero when c > 0 c>0 c > 0 . For every 0 < α < 1 0<\alpha<1 0 < α < 1 ,
C ( c ) , α g a t e = E X P . C^{\mathrm{gate}}_{(c),\alpha}=\mathbf{EXP}. C ( c ) , α gate = EXP . The forward-only kernels D η ( c ) g a t e \mathsf{D}\eta^{\mathrm{gate}}_{(c)} D η ( c ) gate and the kernel obtained from η ( c ) g a t e \eta^{\mathrm{gate}}_{(c)} η ( c ) gate by replacing each step s m → s m − 1 s_m\to s_m-1 s m → s m − 1 with s m → s m + 1 s_m\to s_m+1 s m → s m + 1 have classes contained in P \mathbf{P} P .
Proof. As in the proof of Theorem 6.13(iii), the arrival tick at s m s_m s m is σ m = 3 + ∑ r = 1 m − 2 ( G r + 1 ) \sigma_m=3+\sum_{r=1}^{m-2}(G_r+1) σ m = 3 + ∑ r = 1 m − 2 ( G r + 1 ) , where now the G r G_r G r are independent and G r G_r G r has law Geom ( β r + 1 ) \operatorname{Geom}(\beta_{r+1}) Geom ( β r + 1 ) on { 1 , 2 , … } \{1,2,\ldots\} { 1 , 2 , … } , so E G r = ( r + 1 ) c \mathbb{E}G_r=(r+1)^c E G r = ( r + 1 ) c . Hence, for m ≥ 2 m\ge2 m ≥ 2 ,
E σ m ≤ 3 + ( m − 2 ) ( m c + 1 ) ≤ 3 + 2 m c + 1 ≤ 5 m c + 1 . \mathbb{E}\sigma_m\le3+(m-2)(m^c+1)\le3+2m^{c+1}\le5m^{c+1}. E σ m ≤ 3 + ( m − 2 ) ( m c + 1 ) ≤ 3 + 2 m c + 1 ≤ 5 m c + 1 .
Put e = ⌈ c ⌉ + 1 e=\lceil c\rceil+1 e = ⌈ c ⌉ + 1 . Let a decider run within 2 p ( n ) 2^{p(n)} 2 p ( n ) transitions, where p p p is an integer polynomial with p ≥ 2 p\ge2 p ≥ 2 . Since s p ( n ) ≥ 2 p ( n ) s_{p(n)}\ge2^{p(n)} s p ( n ) ≥ 2 p ( n ) , we have τ 2 p ( n ) ≤ σ p ( n ) \tau_{2^{p(n)}}\le\sigma_{p(n)} τ 2 p ( n ) ≤ σ p ( n ) and E σ p ( n ) ≤ 5 p ( n ) e \mathbb{E}\sigma_{p(n)}\le5p(n)^e E σ p ( n ) ≤ 5 p ( n ) e . Choose an integer k α ≥ 5 / ( 1 − α ) k_\alpha\ge5/(1-\alpha) k α ≥ 5/ ( 1 − α ) and put Q ( n ) = k α p ( n ) e Q(n)=k_\alpha p(n)^e Q ( n ) = k α p ( n ) e . Markov’s inequality gives Pr ( τ 2 p ( n ) > Q ( n ) ) ≤ 1 − α \Pr(\tau_{2^{p(n)}}>Q(n))\le1-\alpha Pr ( τ 2 p ( n ) > Q ( n )) ≤ 1 − α . Crossing times are monotone in the target, and soundness is inherited from the decider, so E X P ⊆ C η ( c ) g a t e , α \mathrm{EXP} \subseteq C_{\eta^{\mathrm{gate}}_{(c)},\alpha} EXP ⊆ C η ( c ) gate , α . Since η ( c ) g a t e ∈ K 1 \eta^{\mathrm{gate}}_{(c)} \in\mathcal{K}_1 η ( c ) gate ∈ K 1 , Theorem 6.13(ii) gives the reverse inclusion, and Theorem 6.13(i) gives the statement about the forward-only kernels.
The return probability may therefore decay polynomially in log 2 i \log_2 i log 2 i along the landing indices. The mechanism is unchanged; only the mean cycle lengths grow, from a constant to O ( m c ) O(m^c) O ( m c ) . At fixed confidence, the crossing budget for T = 2 m T=2^m T = 2 m is O ( m c + 1 ) O(m^{c+1}) O ( m c + 1 ) , still polynomial in log 2 T \log_2 T log 2 T .
Definition 6.17 (Squaring gates) Let Λ 2 = { 2 2 k : k ≥ 1 } = { 4 , 16 , 256 , … } \Lambda_2=\{2^{2^k}: k\ge1\}=\{4,16,256,\ldots\} Λ 2 = { 2 2 k : k ≥ 1 } = { 4 , 16 , 256 , … } . For an integer w ≥ 1 w\ge1 w ≥ 1 , let K w ( 2 ) \mathcal{K}^{(2)}_w K w ( 2 ) be the set of time-homogeneous stochastic kernels η \eta η on N \mathbb{N} N such that η ( g , g 2 + 1 ) = 1 \eta(g,g^2+1)=1 η ( g , g 2 + 1 ) = 1 for every g ∈ Λ 2 g\in\Lambda_2 g ∈ Λ 2 , and, for every i ∉ Λ 2 i\notin\Lambda_2 i ∈ / Λ 2 , η ( i , j ) = 0 \eta(i,j)=0 η ( i , j ) = 0 whenever j > i + w j>i+w j > i + w .
A squaring gate lands one step beyond the next gate g 2 g^2 g 2 . For β ∈ ( 0 , 1 ] \beta\in(0,1] β ∈ ( 0 , 1 ] , define η β s q ∈ K 1 ( 2 ) \eta^{\mathrm{sq}}_\beta\in\mathcal{K}^{(2)}_1 η β sq ∈ K 1 ( 2 ) as follows. Put η β s q ( g , g 2 + 1 ) = 1 \eta^{\mathrm{sq}}_\beta(g,g^2+1)=1 η β sq ( g , g 2 + 1 ) = 1 for g ∈ Λ 2 g\in\Lambda_2 g ∈ Λ 2 . At each landing index s m ′ = 2 2 m + 1 s'_m=2^{2^m}+1 s m ′ = 2 2 m + 1 with m ≥ 2 m\ge2 m ≥ 2 , it moves to s m ′ − 1 s'_m-1 s m ′ − 1 with probability β \beta β and holds with probability 1 − β 1-\beta 1 − β . At every other index it moves to i + 1 i+1 i + 1 . Thus Back ( η β s q ) = { 2 2 m + 1 : m ≥ 2 } \operatorname{Back}(\eta^{\mathrm{sq}}_\beta)=\{2^{2^m}+1:m\ge2\} Back ( η β sq ) = { 2 2 m + 1 : m ≥ 2 } , whose counting function below N N N is at most log 2 log 2 N \log_2\log_2 N log 2 log 2 N for N ≥ 2 N\ge2 N ≥ 2 .
Squaring gates give doubly exponential time using only doubly sparse backward indices.
Theorem 6.18 (Doubly sparse backward indices give 2EXP ) Fix an integer w ≥ 1 w\ge1 w ≥ 1 and 0 < α < 1 0<\alpha<1 0 < α < 1 .
(i) If η ∈ K w ( 2 ) \eta\in\mathcal{K}^{(2)}_w η ∈ K w ( 2 ) is forward-only , then surely max t < q X t ≤ w q + 4 w 2 q 2 \max_{t<q}X_t\le wq+4w^2q^2 max t < q X t ≤ w q + 4 w 2 q 2 , so C η , α ⊆ P C_{\eta,\alpha}\subseteq\mathrm{P} C η , α ⊆ P . Equality holds for the member that moves by + 1 +1 + 1 at every non-gate index.
(ii) Every η ∈ K w ( 2 ) \eta\in\mathcal{K}^{(2)}_w η ∈ K w ( 2 ) satisfies X t ≤ ( 2 w ) 2 t X_t\le(2w)^{2^t} X t ≤ ( 2 w ) 2 t surely, so C η , α ⊆ 2 E X P C_{\eta,\alpha}\subseteq\mathrm{2EXP} C η , α ⊆ 2EXP .
(iii) For every β ∈ ( 0 , 1 ] \beta\in(0,1] β ∈ ( 0 , 1 ] , C η β s q , α = 2 E X P C_{\eta^{\mathrm{sq}}_\beta,\alpha}=\mathrm{2EXP} C η β sq , α = 2EXP , and the crossing budget of η β s q \eta^{\mathrm{sq}}_\beta η β sq satisfies
q α ( T ) log 2 log 2 T ⟶ 1 + β − 1 ( T → ∞ ) . \frac{q_\alpha(T)}{\log_2\log_2 T}\longrightarrow1+\beta^{-1}\qquad(T\to\infty). log 2 log 2 T q α ( T ) ⟶ 1 + β − 1 ( T → ∞ ) .
Proof. (i) Argue as in Theorem 6.13(i). A forward-only path is nondecreasing, and each gate is occupied at no more than one tick. List the gates occupied before tick t t t as g 1 < ⋯ < g N g_1<\cdots<g_N g 1 < ⋯ < g N , put λ 0 = 0 \lambda_0=0 λ 0 = 0 and λ r = g r 2 + 1 \lambda_r=g_r^2+1 λ r = g r 2 + 1 , and let ω r = g r − λ r − 1 \omega_r=g_r-\lambda_{r-1} ω r = g r − λ r − 1 be the total non-gate displacement between the departure from g r − 1 g_{r-1} g r − 1 and the arrival at g r g_r g r , starting at tick zero when r = 1 r=1 r = 1 . For r = 1 r=1 r = 1 , ω 1 = g 1 \omega_1=g_1 ω 1 = g 1 . For r ≥ 2 r\ge2 r ≥ 2 , write g r − 1 = 2 2 j g_{r-1}=2^{2^j} g r − 1 = 2 2 j . Since g r > λ r − 1 = 2 2 j + 1 + 1 g_r>\lambda_{r-1}=2^{2^{j+1}}+1 g r > λ r − 1 = 2 2 j + 1 + 1 , we have g r ≥ 2 2 j + 2 g_r\ge2^{2^{j+2}} g r ≥ 2 2 j + 2 , so λ r − 1 ≤ g r + 1 \lambda_{r-1}\le\sqrt{g_r}+1 λ r − 1 ≤ g r + 1 and ω r ≥ g r − g r − 1 ≥ g r / 2 \omega_r\ge g_r-\sqrt{g_r}-1\ge g_r/2 ω r ≥ g r − g r − 1 ≥ g r /2 because g r ≥ 256 g_r\ge256 g r ≥ 256 . In both cases g r ≤ 2 ω r g_r\le2\omega_r g r ≤ 2 ω r , so the gain at the r r r th gate satisfies g r 2 + 1 − g r ≤ g r 2 ≤ 4 ω r 2 g_r^2+1-g_r\le g_r^2\le4\omega_r^2 g r 2 + 1 − g r ≤ g r 2 ≤ 4 ω r 2 . With Ω t ≤ w t \Omega_t\le wt Ω t ≤ w t the total non-gate displacement before tick t t t , the disjointness of the stretches gives
X t ≤ Ω t + 4 ∑ r ω r 2 ≤ Ω t + 4 Ω t 2 ≤ w t + 4 w 2 t 2 . X_t\le\Omega_t+4\sum_r\omega_r^2\le\Omega_t+4\Omega_t^2\le wt+4w^2t^2. X t ≤ Ω t + 4 r ∑ ω r 2 ≤ Ω t + 4 Ω t 2 ≤ w t + 4 w 2 t 2 .
The path is nondecreasing, so this bounds the running maximum. A witness with budget Q Q Q therefore accepts at indices at most w Q ( n ) + 4 w 2 Q ( n ) 2 wQ(n)+4w^2Q(n)^2 w Q ( n ) + 4 w 2 Q ( n ) 2 , and simulation gives a polynomial-time decider. For the member moving by + 1 +1 + 1 at non-gate indices, X t ≥ t X_t\ge t X t ≥ t surely, which gives P ⊆ C η , α \mathrm{P}\subseteq C_{\eta,\alpha} P ⊆ C η , α .
(ii) Put Y t = X t + w Y_t=X_t+w Y t = X t + w . At a gate, X t + 1 = X t 2 + 1 ≤ Y t 2 X_{t+1}=X_t^2+1\le Y_t^2 X t + 1 = X t 2 + 1 ≤ Y t 2 ; elsewhere X t + 1 ≤ X t + w = Y t ≤ Y t 2 X_{t+1}\le X_t+w=Y_t\le Y_t^2 X t + 1 ≤ X t + w = Y t ≤ Y t 2 . Hence Y t + 1 ≤ Y t 2 + w ≤ 2 Y t 2 Y_{t+1}\le Y_t^2+w\le2Y_t^2 Y t + 1 ≤ Y t 2 + w ≤ 2 Y t 2 , so 2 Y t + 1 ≤ ( 2 Y t ) 2 2Y_{t+1}\le(2Y_t)^2 2 Y t + 1 ≤ ( 2 Y t ) 2 . Since Y 0 = w Y_0=w Y 0 = w , induction gives 2 Y t ≤ ( 2 w ) 2 t 2Y_t\le(2w)^{2^t} 2 Y t ≤ ( 2 w ) 2 t . Thus X t ≤ ( 2 w ) 2 t X_t\le(2w)^{2^t} X t ≤ ( 2 w ) 2 t . A witness with budget Q Q Q accepts at indices at most ( 2 w ) 2 Q ( n ) (2w)^{2^{Q(n)}} ( 2 w ) 2 Q ( n ) , and simulating through that bound is a doubly exponential-time decider.
(iii) Write d β = 1 + β − 1 d_\beta=1+\beta^{-1} d β = 1 + β − 1 . From zero the path visits 0 , 1 , 2 , 3 , 4 0,1,2,3,4 0 , 1 , 2 , 3 , 4 at ticks 0 , … , 4 0,\ldots,4 0 , … , 4 and s 2 ′ = 17 s'_2=17 s 2 ′ = 17 at tick 5. For r ≥ 1 r\ge1 r ≥ 1 , the clock waits G r G_r G r ticks at s r + 1 ′ s'_{r+1} s r + 1 ′ until it steps to the gate 2 2 r + 1 2^{2^{r+1}} 2 2 r + 1 , and the next tick moves it to s r + 2 ′ s'_{r+2} s r + 2 ′ . The G r G_r G r are iid Geom ( β ) \operatorname{Geom}(\beta) Geom ( β ) on { 1 , 2 , … } \{1,2,\ldots\} { 1 , 2 , … } , so the arrival tick at s m ′ s'_m s m ′ is
σ m ′ = 5 + ∑ r = 1 m − 2 ( G r + 1 ) ( m ≥ 2 ) . \sigma'_m = 5 + \sum_{r=1}^{m-2}(G_r+1) \qquad(m \ge2). σ m ′ = 5 + r = 1 ∑ m − 2 ( G r + 1 ) ( m ≥ 2 ) .
For m ≥ 3 m \ge3 m ≥ 3 , before σ m ′ \sigma'_m σ m ′ every visited index is at most s m − 1 ′ s'_{m-1} s m − 1 ′ . For an integer T ≥ 18 T \ge18 T ≥ 18 , put m ( T ) = ⌈ log 2 log 2 ( T − 1 ) ⌉ ≥ 3 m(T)=\left\lceil\log_2\log_2(T-1)\right\rceil\ge3 m ( T ) = ⌈ log 2 log 2 ( T − 1 ) ⌉ ≥ 3 . Then s m ( T ) − 1 ′ < T ≤ s m ( T ) ′ s'_{m(T)-1}<T\le s'_{m(T)} s m ( T ) − 1 ′ < T ≤ s m ( T ) ′ , so τ T = σ m ( T ) ′ \tau_T=\sigma'_{m(T)} τ T = σ m ( T ) ′ . The strong law of large numbers and m ( T ) / log 2 log 2 T → 1 m(T)/\log_2\log_2 T\to1 m ( T ) / log 2 log 2 T → 1 give τ T / log 2 log 2 T → d β \tau_T/\log_2\log_2 T\to d_\beta τ T / log 2 log 2 T → d β almost surely. The quantile limit follows as in Theorem 6.13(iii).
For 2 E X P ⊆ C η s q , α 2\mathrm{EXP}\subseteq\mathcal{C}_{\eta}^{\mathrm{sq},\alpha} 2 EXP ⊆ C η sq , α , let a decider run within 2 2 p ( n ) 2^{2^{p(n)}} 2 2 p ( n ) transitions, where p p p is an integer polynomial with p ≥ 2 p\ge2 p ≥ 2 . Since s p ( n ) ′ ≥ 2 2 p ( n ) s'_{p(n)}\ge2^{2^{p(n)}} s p ( n ) ′ ≥ 2 2 p ( n ) , we have τ 2 p ( n ) ≤ σ p ( n ) ′ \tau_{2^{p(n)}}\le\sigma'_{p(n)} τ 2 p ( n ) ≤ σ p ( n ) ′ , and E σ p ( n ) ′ ≤ 5 + d β p ( n ) ≤ ( 5 + d β ) p ( n ) \mathbb{E}\sigma'_{p(n)}\le5+d_\beta p(n)\le(5+d_\beta)p(n) E σ p ( n ) ′ ≤ 5 + d β p ( n ) ≤ ( 5 + d β ) p ( n ) . With an integer k ≥ ( 5 + d β ) / ( 1 − α ) k\ge(5+d_\beta)/(1-\alpha) k ≥ ( 5 + d β ) / ( 1 − α ) and Q ( n ) = k p ( n ) Q(n)=kp(n) Q ( n ) = k p ( n ) , Markov’s inequality gives completeness at confidence α \alpha α . Soundness is inherited from the decider. Part (ii) gives the reverse inclusion. □
Fixed iid laws also reach 2 E X P 2\mathrm{EXP} 2 EXP (Theorem 5.8). Theorem 6.18 concerns the comparison inside K w ( 2 ) \mathcal{K}_{w}^{(2)} K w ( 2 ) , where forward-only members stay in P \mathbf{P} P .
Remark 6.19 (How sparse can backward indices be?) In both gated families the backward indices are the landing points, so their sparsity is set by the spacing of the gates. Forward-only members stay in P \mathbf{P} P because the gain at a gate is polynomial in the walk needed to reach it: linear for doubling gates, quadratic for squaring gates. These constructions give counting orders log N \log N log N and log log N \log\log N log log N , respectively. A gate g ↦ 2 g g\mapsto2^g g ↦ 2 g would give a forward-only clock an exponential gain after a walk of order g g g , outside this polynomial-gain argument. We leave open whether backward indices of arbitrarily small counting order can raise the class beyond P \mathbf{P} P in a family whose forward-only members stay in P \mathbf{P} P .
Conclusions and open problems
Crossing reach connects a fixed random clock to its observation class. The sparse computable constructions give one zero-drift intermediate class and an infinite negative-drift antichain below P \mathrm{P} P . The gated constructions give a different comparison: under their prescribed gate jumps and bounded forward displacement away from gates, every forward-only member stays within P \mathrm{P} P , whereas length-one returns attain E X P \mathrm{EXP} EXP or 2 E X P \mathrm{2EXP} 2EXP . The parity example establishes a gain over deletion, with forward re-entry also attaining E X P \mathrm{EXP} EXP . In both gated families, sparsity concerns backward-index support in index space. The doubling family also permits vanishing return probabilities at its specified landing indices. These classifications use the unit-cost observation convention.
The classifications leave six questions. Each concerns an unresolved regime or the boundary of the present constructions.
Determine the exact class of the sparse zero-drift law of Theorem 4.6 .
Classify negative-mean laws with only some finite positive moments, including the antichain regime of Theorem 4.8 .
Classify integrable zero-drift laws with infinite variance and unbounded positive jumps.
Determine how confidence affects iid classes beyond the classified cases. Appendix § 13 .§ 7 gives a position-dependent example; Appendix § 13 .8 transfers iid bounds but leaves the sparse case open.
Determine the sparsest backward-index sets that can raise a gated family beyond P \mathrm{P} P (Theorem 6.19).
Determine whether the law of Theorem D.1 is computable.
Random input access or charging configuration generation would require a new analysis of the cost model.
Statements and Declarations
Funding. No specific funding was received for this research.
Competing interests. The authors declare no competing interests.
Data availability. This article is theoretical. No empirical datasets were generated or analysed. The mathematical constructions and proofs are provided in the article.
Use of LLM. The authors occasionally used LLM under their guidance to assist with developing the proofs, editing, and formatting. The authors critically reviewed, revised, and verified the final content and take full responsibility for the manuscript.
The computable zero-drift construction
We prove Theorem 4.6 by placing sparse atoms and balancing the mean. At chosen even targets T j T_j T j , the budgets q j = T j j q_j = T_j^j q j = T j j eventually dominate every fixed polynomial evaluated at T j T_j T j , yet the crossing probability is below 2 − j 2^{-j} 2 − j . Finite searches make the law computable, and recursive choices preserve all earlier bounds.
Proof of Theorem 4.6. Subscripts on probabilities and expectations in this proof specify the proposal law. Put a j = 2 − j − 1 a_j = 2^{-j-1} a j = 2 − j − 1 for j ≥ 1 j \ge1 j ≥ 1 . Then ∑ j ≥ 1 a j = 1 / 2 \sum_{j\ge1} a_j = 1/2 ∑ j ≥ 1 a j = 1/2 and ∑ i ≥ j a i = 2 − j \sum_{i\ge j} a_i = 2^{-j} ∑ i ≥ j a i = 2 − j . We will choose N j N_j N j recursively and set
μ ( − 1 ) = 1 2 , μ ( N j ) = a j N j , μ ( 0 ) = 1 2 − ∑ j ≥ 1 a j N j . (25) \mu(-1) = \frac{1}{2}, \qquad\mu(N_j) = \frac{a_j}{N_j}, \qquad\mu(0) = \frac{1}{2} - \sum_{j\ge1} \frac{a_j}{N_j}. \tag*{(25)} μ ( − 1 ) = 2 1 , μ ( N j ) = N j a j , μ ( 0 ) = 2 1 − j ≥ 1 ∑ N j a j . (25)
with zero mass elsewhere. Since N j ≥ 1 N_j \ge1 N j ≥ 1 , the series is at most 1 / 2 1/2 1/2 . The zero mass is therefore nonnegative and the total mass is one, with
E K + = ∑ j ≥ 1 a j = 1 2 = E K − , E ∣ K ∣ = 1 , E K = 0. \mathbb{E}K^+ = \sum_{j\ge1} a_j = \frac{1}{2} = \mathbb{E}K^-, \qquad\mathbb{E}|K| = 1, \qquad\mathbb{E}K = 0. E K + = j ≥ 1 ∑ a j = 2 1 = E K − , E ∣ K ∣ = 1 , E K = 0.
Effective finite approximations. Set N 0 = T 0 = 0 N_0 = T_0 = 0 N 0 = T 0 = 0 . At stage j j j , only N 1 , … , N j − 1 N_1,\ldots,N_{j-1} N 1 , … , N j − 1 have been chosen. Define the finite rational law ν j \nu_j ν j below, with all atom indices in the sums starting at one:
ν j ( − 1 ) = 1 2 , ν j ( N i ) = a i N i ( i < j ) , ν j ( 0 ) = 1 2 − ∑ i < j a i N i . \nu_j(-1) = \frac{1}{2}, \qquad\nu_j(N_i) = \frac{a_i}{N_i}\ (i<j), \qquad\nu_j(0) = \frac{1}{2} - \sum_{i<j} \frac{a_i}{N_i}. ν j ( − 1 ) = 2 1 , ν j ( N i ) = N i a i ( i < j ) , ν j ( 0 ) = 2 1 − i < j ∑ N i a i .
It has drift − 1 / 2 + ∑ i < j a i = − 2 − j < 0 -1/2 + \sum_{i<j} a_i = -2^{-j} < 0 − 1/2 + ∑ i < j a i = − 2 − j < 0 . In the completed law, this is the distribution obtained by replacing every proposal N i N_i N i with i ≥ j i \ge j i ≥ j by zero.
Search over m = 1 , 2 , … m = 1,2,\ldots m = 1 , 2 , … for a rational r j = 1 + 2 − m r_j = 1 + 2^{-m} r j = 1 + 2 − m such that
λ j = E ν j [ r j K ] < 1. \lambda_j = \mathbb{E}_{\nu_j}[r_j^K] < 1. λ j = E ν j [ r j K ] < 1.
Every test uses exact rational arithmetic on a finite sum. The derivative of r ↦ E ν j [ r K ] r \mapsto\mathbb{E}_{\nu_j}[r^K] r ↦ E ν j [ r K ] at r = 1 r = 1 r = 1 is − 2 − j -2^{-j} − 2 − j , so the sum is less than one for all r > 1 r > 1 r > 1 sufficiently close to one. The search therefore terminates.
For V ( i ) = r j i V(i) = r_j^i V ( i ) = r j i , at every i ≥ 1 i \ge1 i ≥ 1 all proposals of ν j \nu_j ν j are admissible, since its only negative proposal is − 1 -1 − 1 . At zero that proposal holds. Hence, with b j = ( 1 − r j − 1 ) / 2 b_j = (1-r_j^{-1})/2 b j = ( 1 − r j − 1 ) /2 , the exact identity is
E ν j [ V ( X t + 1 ) ∣ X t = i ] = λ j V ( i ) + b j 1 { i = 0 } ≤ λ j V ( i ) + b j . \mathbb{E}_{\nu_j}[V(X_{t+1}) \mid X_t = i] = \lambda_j V(i) + b_j \mathbf{1}_{\{i=0\}} \le\lambda_j V(i) + b_j. E ν j [ V ( X t + 1 ) ∣ X t = i ] = λ j V ( i ) + b j 1 { i = 0 } ≤ λ j V ( i ) + b j .
Iteration from X 0 = 0 X_0 = 0 X 0 = 0 gives the computable rational bound sup t E ν j V ( X t ) ≤ C j : = 1 + b j / ( 1 − λ j ) \sup_t \mathbb{E}_{\nu_j}V(X_t) \le C_j := 1+b_j/(1-\lambda_j) sup t E ν j V ( X t ) ≤ C j := 1 + b j / ( 1 − λ j ) . Markov’s inequality and Boole’s inequality (the union bound) [18 Theorem 1.1.1] over ticks yield
Pr ν j ( τ T ≤ q ) ≤ C j ( q + 1 ) r j − T ( T ≥ 1 , q ≥ 0 ) . (26) \Pr_{\nu_j}(\tau_T \le q) \le C_j(q+1)r_j^{-T} \qquad(T \ge1,\ q \ge0). \tag*{(26)} ν j Pr ( τ T ≤ q ) ≤ C j ( q + 1 ) r j − T ( T ≥ 1 , q ≥ 0 ) . (26)
Choosing the targets and atoms. Search in increasing order for an even integer T j ≥ max ( 2 , j ) T_j \ge\max(2,j) T j ≥ max ( 2 , j ) with T j > T j − 1 T_j > T_{j-1} T j > T j − 1 and
C j ( T j j + 1 ) r j − T j < 2 − j − 1 . (27) C_j(T_j^j+1)r_j^{-T_j} < 2^{-j-1}. \tag*{(27)} C j ( T j j + 1 ) r j − T j < 2 − j − 1 . (27)
For fixed j j j and r j > 1 r_j > 1 r j > 1 , ( T j j + 1 ) r j − T j → 0 (T_j^j+1)r_j^{-T_j} \to0 ( T j j + 1 ) r j − T j → 0 along the even integers, so this exact rational search terminates. Put q j = T j j q_j = T_j^j q j = T j j and choose the least integer
N j > max { N j − 1 , 2 q j , 2 2 j } . (28) N_j > \max\{N_{j-1}, 2q_j, 2^{2j}\}. \tag*{(28)} N j > max { N j − 1 , 2 q j , 2 2 j } . (28)
These rules define computable increasing sequences independently of the machine, input, polynomial budget, and confidence.
The completed law. For every j j j , monotonicity of the atoms and the sum of the a i a_i a i give
∑ i ≥ j μ ( N i ) = ∑ i ≥ j a i N i ≤ 2 − j N j . \sum_{i\ge j}\mu(N_i)=\sum_{i\ge j}\frac{a_i}{N_i}\le\frac{2^{-j}}{N_j}. i ≥ j ∑ μ ( N i ) = i ≥ j ∑ N i a i ≤ N j 2 − j .
Couple a walk with law μ \mu μ to one with law ν j \nu_j ν j by replacing a proposal N i N_i N i , i ≥ j i\ge j i ≥ j , by zero and leaving all other proposals unchanged. The transformed sequence has independent proposals with law ν j \nu_j ν j . The two held walks agree until the first replacement. A union bound over the first q j q_j q j proposals, followed by (A2)–(A4), gives
Pr μ ( τ T j ≤ q j ) ≤ C j ( q j + 1 ) r j − T j + q j ∑ i ≥ j μ ( N i ) < 2 − j − 1 + 2 − j − 1 = 2 − j . (29) \Pr_{\mu}(\tau_{T_j}\le q_j)\le C_j(q_j+1)r_j^{-T_j}+q_j\sum_{i\ge j}\mu(N_i)<2^{-j-1}+2^{-j-1}=2^{-j}. \tag*{(29)} μ Pr ( τ T j ≤ q j ) ≤ C j ( q j + 1 ) r j − T j + q j i ≥ j ∑ μ ( N i ) < 2 − j − 1 + 2 − j − 1 = 2 − j . (29)
The sum includes every future atom of the infinite law. Later choices therefore preserve this bound at every earlier stage.
Even length and the quantifiers. Fix any 0 < α < 1 0<\alpha<1 0 < α < 1 and suppose ( M , Q ) (M,Q) ( M , Q ) were a witness for L e v e n L_{\mathrm{even}} L even . Let x x x be an accepted word of positive even length n n n , and choose a ∈ Σ a\in\Sigma a ∈ Σ . The tapes for x x x and x a xa x a initially differ only at cell n + 1 n+1 n + 1 : it is blank for x x x and contains a a a for x a xa x a . Until that cell is read, deterministic transitions, including any writes to earlier cells, are identical on the two inputs. Accepting x x x before reading cell n + 1 n+1 n + 1 would therefore accept the odd-length word x a xa x a , contrary to soundness. The t t t th transition can read at most cell t t t , so reading cell n + 1 n+1 n + 1 requires at least n + 1 n+1 n + 1 transitions. Therefore A M ( x ) ≥ n + 1 A_M(x)\ge n+1 A M ( x ) ≥ n + 1 ; the construction only needs A M ( x ) ≥ n A_M(x)\ge n A M ( x ) ≥ n . We use only positive even lengths below; ε \varepsilon ε needs no lower bound.
Choose integers c ≥ 1 c\ge1 c ≥ 1 and d ≥ 0 d\ge0 d ≥ 0 with Q ( n ) ≤ c n d Q(n)\le cn^d Q ( n ) ≤ c n d for n ≥ 1 n\ge1 n ≥ 1 ; the sum of the coefficients, enlarged to one, suffices for c c c . For all sufficiently large j j j , j ≥ d j\ge d j ≥ d and 2 j − d ≥ c 2^{j-d}\ge c 2 j − d ≥ c . Since T j ≥ 2 T_j\ge2 T j ≥ 2 , this implies Q ( T j ) ≤ c T j d ≤ T j j = q j Q(T_j)\le cT_j^d\le T_j^j=q_j Q ( T j ) ≤ c T j d ≤ T j j = q j . The same conclusion holds for any polynomial upper bound after enlarging it to an integer polynomial as in Section 2. For the yes-input x j = a T j x_j=a^{T_j} x j = a T j , monotonicity in the target and budget now gives
Pr μ ( τ A M ( x j ) ≤ Q ( T j ) ) ≤ Pr μ ( τ T j ≤ q j ) < 2 − j . \Pr_{\mu}(\tau_{A_M(x_j)}\le Q(T_j))\le\Pr_{\mu}(\tau_{T_j}\le q_j)<2^{-j}. μ Pr ( τ A M ( x j ) ≤ Q ( T j )) ≤ μ Pr ( τ T j ≤ q j ) < 2 − j .
Taking j j j large enough that 2 − j < α 2^{-j}<\alpha 2 − j < α contradicts completeness. This excludes every witness at every fixed confidence for the same law. The finite first positive moment gives containment in P \mathbf{P} P by Theorem 3.1 ; L e v e n L_{\mathrm{even}} L even makes the containment strict. The universal lower bound of Theorem 2.17 gives R D E F ⊆ P μ , α ∗ \mathbf{RDEF}\subseteq\mathbf{P}^{*}_{\mu,\alpha} RDEF ⊆ P μ , α ∗ , since Pr ( K > 0 ) > 0 \Pr(K>0)>0 Pr ( K > 0 ) > 0 .
Computability and variance. Each stage above terminates by a proved strict-inequality search over finite rational data. To compute μ ( k ) \mu(k) μ ( k ) for a positive integer k k k , generate the N j N_j N j until the first N j ≥ k N_j\ge k N j ≥ k . If equality holds, return a j / N j a_j/N_j a j / N j ; otherwise return zero. This terminates because N j > 2 2 j N_j>2^{2^j} N j > 2 2 j . For k = − 1 k=-1 k = − 1 return 1 / 2 1/2 1/2 , and for k < − 1 k<-1 k < − 1 return zero. At zero, after computing the first s s s atoms, the remaining mass satisfies
0 < ∑ i > s a i N i ≤ ∑ i > s a i = 2 − s − 1 . 0<\sum_{i>s}\frac{a_i}{N_i}\le\sum_{i>s}a_i=2^{-s-1}. 0 < i > s ∑ N i a i ≤ i > s ∑ a i = 2 − s − 1 .
Thus 1 / 2 − ∑ i = 1 s a i / N i 1/2-\sum_{i=1}^{s}a_i/N_i 1/2 − ∑ i = 1 s a i / N i approximates μ ( 0 ) \mu(0) μ ( 0 ) with error at most 2 − s − 1 2^{-s-1} 2 − s − 1 . Given a requested error 2 − p 2^{-p} 2 − p , choose s ≥ p s\ge p s ≥ p . This gives uniform computability of the point probabilities. Finally
∑ j ≥ 1 N j 2 μ ( N j ) = ∑ j ≥ 1 a j N j = ∞ , a j N j > 2 j − 1 . \sum_{j\ge1}N_j^2\mu(N_j)=\sum_{j\ge1}a_jN_j=\infty,\qquad a_jN_j>2^{j-1}. j ≥ 1 ∑ N j 2 μ ( N j ) = j ≥ 1 ∑ a j N j = ∞ , a j N j > 2 j − 1 .
The mean is zero, so this infinite second moment is also infinite variance.
A strictly intermediate class
The same law admits a language outside RDEF . We use input lengths large enough to pay for regenerating an atom, but small enough that a polynomial observation budget crosses that atom.
Strict lower inclusion in Theorem 4.6 . Fix a finite nonempty alphabet Σ \Sigma Σ . The law and the sequences ( T j ) (T_j) ( T j ) and ( N j ) (N_j) ( N j ) above do not depend on Σ \Sigma Σ . Since μ ( N j ) = 2 − j − 1 / N j \mu(N_j)=2^{-j-1}/N_j μ ( N j ) = 2 − j − 1 / N j and N j > 2 2 j N_j>2^{2j} N j > 2 2 j ,
2 j + 1 < 2 N j 1 / 2 , μ ( N j ) > 1 2 N j 3 / 2 . 2^{j+1}<2N_j^{1/2},\qquad\mu(N_j)>\frac{1}{2N_j^{3/2}}. 2 j + 1 < 2 N j 1/2 , μ ( N j ) > 2 N j 3/2 1 .
A proposal N j N_j N j crosses threshold N j N_j N j from every nonnegative position. For any fixed 0 < α < 1 0<\alpha<1 0 < α < 1 , independence therefore gives
Pr μ ( τ N j ≤ q ) ≥ 1 − ( 1 − μ ( N j ) ) q ≥ 1 − exp ( − q μ ( N j ) ) ≥ α if q ≥ 2 ℓ α N j 3 / 2 . (30) \Pr_{\mu}(\tau_{N_j}\le q)\ge1-(1-\mu(N_j))^q\ge1-\exp(-q\mu(N_j))\ge\alpha\quad\text{if }q\ge2\ell_{\alpha}N_j^{3/2}. \tag*{(30)} μ Pr ( τ N j ≤ q ) ≥ 1 − ( 1 − μ ( N j ) ) q ≥ 1 − exp ( − q μ ( N j )) ≥ α if q ≥ 2 ℓ α N j 3/2 . (30)
Thus R α ( q ) ≥ N j R_{\alpha}(q)\ge N_j R α ( q ) ≥ N j at these integer budgets.
We next bound the cost of the exact searches, including a single-tape implementation. For j ≥ 2 j\ge2 j ≥ 2 , put N = N j − 1 ≥ 1 N=N_{j-1}\ge1 N = N j − 1 ≥ 1 and let m j m_j m j be the first successful exponent in the search for r j = 1 + 2 − m j r_j=1+2^{-m_j} r j = 1 + 2 − m j . If 0 < ε ≤ 1 / N 0<\varepsilon\le1/N 0 < ε ≤ 1/ N , Taylor’s formula gives ( 1 + ε ) k ≤ 1 + k ε + ( e / 2 ) k 2 ε 2 (1+\varepsilon)^k\le1+k\varepsilon+(e/2)k^2\varepsilon^2 ( 1 + ε ) k ≤ 1 + k ε + ( e /2 ) k 2 ε 2 for 1 ≤ k ≤ N 1\le k\le N 1 ≤ k ≤ N . Also ( 1 + ε ) − 1 ≤ 1 − ε + ε 2 (1+\varepsilon)^{-1}\le1-\varepsilon+\varepsilon^2 ( 1 + ε ) − 1 ≤ 1 − ε + ε 2 . Since ∑ i < j a i = 1 / 2 − 2 − j \sum_{i<j}a_i=1/2-2^{-j} ∑ i < j a i = 1/2 − 2 − j and ∑ i < j a i N i ≤ N / 2 \sum_{i<j}a_iN_i\le N/2 ∑ i < j a i N i ≤ N /2 , we obtain
E ν j ( 1 + ε ) K − 1 ≤ − 2 − j ε + ( 1 2 + e 4 N ) ε 2 ≤ − 2 − j ε + 2 N ε 2 . \mathbb{E}_{\nu_j}(1+\varepsilon)^K-1\le-2^{-j}\varepsilon+\left(\frac{1}{2}+\frac{e}{4}N\right)\varepsilon^2\le-2^{-j}\varepsilon+2N\varepsilon^2. E ν j ( 1 + ε ) K − 1 ≤ − 2 − j ε + ( 2 1 + 4 e N ) ε 2 ≤ − 2 − j ε + 2 N ε 2 .
Taking ε ≤ 2 − j / ( 4 N ) \varepsilon\le2^{-j}/(4N) ε ≤ 2 − j / ( 4 N ) makes this negative. Consequently
m j ≤ j + ⌈ log 2 N ⌉ + 2. (31) m_j\le j+\lceil\log_2N\rceil+2. \tag*{(31)} m j ≤ j + ⌈ log 2 N ⌉ + 2. (31)
Stage one has fixed finite cost.
The earlier atom has mass a j − 1 / N = 2 − j / N a_{j-1}/N=2^{-j}/N a j − 1 / N = 2 − j / N in ν j \nu_j ν j . Every term in λ j \lambda_j λ j is nonnegative, so λ j < 1 \lambda_j<1 λ j < 1 implies r j N < 2 j N r_j^N<2^jN r j N < 2 j N . On the other hand, (A3) and C j ≥ 1 C_j\ge1 C j ≥ 1 imply r j T j > 2 j + 1 r_j^{T_j}>2^{j+1} r j T j > 2 j + 1 . Taking natural logarithms yields
T j > ( j + 1 ) ln 2 ln r j > ( j + 1 ) ( ln 2 ) N j ln 2 + ln N . (32) T_j>\frac{(j+1)\ln2}{\ln r_j}>\frac{(j+1)(\ln2)N}{j\ln2+\ln N}. \tag*{(32)} T j > ln r j ( j + 1 ) ln 2 > j ln 2 + ln N ( j + 1 ) ( ln 2 ) N . (32)
By (A4), N > 2 2 ( j − 1 ) N>2^{2(j-1)} N > 2 2 ( j − 1 ) , so j = O ( log N ) j=O(\log N) j = O ( log N ) . In particular T j ≥ c N / log N T_j\ge cN/\log N T j ≥ c N / log N for an absolute c > 0 c>0 c > 0 at large stages. This also gives T j ≥ N T_j\ge\sqrt{N} T j ≥ N eventually, and then log N ≤ 2 log T j \log N\le2\log T_j log N ≤ 2 log T j . It follows that
N j − 1 = O ( T j log T j ) , j + m j = O ( log T j ) . (33) N_{j-1}=O(T_j\log T_j),\qquad j+m_j=O(\log T_j). \tag*{(33)} N j − 1 = O ( T j log T j ) , j + m j = O ( log T j ) . (33)
All constants are independent of the stage.
Here is a bit-cost bound for the searches. For a tested exponent m ≤ m j m\le m_j m ≤ m j , the powers ( 1 + 2 − m ) N i (1+2^{-m})^{N_i} ( 1 + 2 − m ) N i can be calculated by binary exponentiation with exact integer numerators and denominators. A common denominator for the finite expectation has bit length O ( m N + j log ( N + 1 ) + j ) O(mN+j\log(N+1)+j) O ( m N + j log ( N + 1 ) + j ) : one can use the common dyadic power 2 m N 2^{mN} 2 m N , the product of the earlier N i N_i N i , a dyadic denominator for the a i a_i a i , and 2 m + 1 2^m+1 2 m + 1 for the negative atom. The rational C j C_j C j has the same order of bit length. At a tested target T ≤ T j T\le T_j T ≤ T j , the comparison in (A3) uses integers with at most
B j = O ( m j ( N + T j ) + j log ( N + T j + 1 ) ) = O ( T j ( log T j ) 2 ) B_j=O\left(m_j(N+T_j)+j\log(N+T_j+1)\right)=O\left(T_j(\log T_j)^2\right) B j = O ( m j ( N + T j ) + j log ( N + T j + 1 ) ) = O ( T j ( log T j ) 2 )
bits. This includes T j T^j T j , the rational powers, and the exact strict comparison; no numerical precision search is needed. There are at most m j m_j m j exponent tests and T j T_j T j target tests, each using at most j j j terms and binary exponentiation. Schoolbook integer arithmetic bounds these operations by a polynomial in N , T j , j , m j N,T_j,j,m_j N , T j , j , m j of a degree independent of j j j . Finally setting N j N_j N j by (A4) uses O ( j log T j + log ( N + 1 ) + j ) O(j\log T_j+\log(N+1)+j) O ( j log T j + log ( N + 1 ) + j ) bits. By (A9), the entire stage costs at most C ( T j + 1 ) D C(T_j+1)^D C ( T j + 1 ) D for fixed C , D C,D C , D . Polynomial single-tape simulation overhead can be included by increasing D D D [16 ] .
Choose an integer j 0 j_0 j 0 sufficiently large, as specified below, and put
n j = ⌊ N j 1 / 2 ⌋ , L G = { x ∈ Σ ∗ : ∣ x ∣ = n j for some j ≥ j 0 } . (34) n_j=\left\lfloor N_j^{1/2}\right\rfloor,\qquad L_G=\{x\in\Sigma^*:|x|=n_j\text{ for some }j\ge j_0\}. \tag*{(34)} n j = ⌊ N j 1/2 ⌋ , L G = { x ∈ Σ ∗ : ∣ x ∣ = n j for some j ≥ j 0 } . (34)
A machine first computes n = ∣ x ∣ n=|x| n = ∣ x ∣ and then regenerates the stages by the exact searches above. At stage i i i it computes ⌊ N i ⌋ \left\lfloor\sqrt{N_i}\right\rfloor ⌊ N i ⌋ by integer arithmetic. It accepts if this equals n n n and i ≥ j 0 i\ge j_0 i ≥ j 0 , and rejects as soon as it exceeds n n n . The search terminates since the N i N_i N i increase without bound. This machine accepts precisely L G L_G L G .
For completeness, input-length computation need not incur quadratic single-tape cost. Mark all input cells on an auxiliary symbol track, and make repeated sweeps that retain every second surviving mark. The parity of the number of marks before each sweep records the next binary digit of n n n on a separate track near the left end. There are O ( log ( n + 2 ) ) O(\log(n+2)) O ( log ( n + 2 )) sweeps, each costing O ( n + 1 ) O(n+1) O ( n + 1 ) transitions. After recording n n n , erase the other input tracks in one sweep and return to the left end. This takes O ( ( n + 1 ) log ( n + 2 ) ) O((n+1)\log(n+2)) O (( n + 1 ) log ( n + 2 )) transitions and leaves the binary length in a local workspace. The subsequent arithmetic is simulated in that workspace, so its polynomial simulation cost does not square the input scan.
On a yes-input of length n j n_j n j , all stage comparisons with n j n_j n j can be included in a fixed polynomial in T j T_j T j . Indeed log N j = O ( j log T j + log ( N j − 1 + 1 ) + j ) \log N_j=O(j\log T_j+\log(N_{j-1}+1)+j) log N j = O ( j log T j + log ( N j − 1 + 1 ) + j ) by the least-integer choice in (A4). This bounds both the length of n j n_j n j and the cost of its comparisons and of the integer square roots. Since T i ≤ T j T_i\le T_j T i ≤ T j for i ≤ j i\le j i ≤ j and j ≤ T j j\le T_j j ≤ T j , there are constants C ′ , D ′ C',D' C ′ , D ′ such that the total single-tape arithmetic time through stage j j j is at most C ′ T j D ′ C'T_j^{D'} C ′ T j D ′ . But N j > 2 T j j N_j>2T_j^j N j > 2 T j j , so
C ′ T j D ′ N j < C ′ 2 T j D ′ − j ⟶ 0. \frac{C'T_j^{D'}}{N_j}<\frac{C'}{2}T_j^{D'-j}\longrightarrow0. N j C ′ T j D ′ < 2 C ′ T j D ′ − j ⟶ 0.
The length-computation cost is also o ( N j ) o(N_j) o ( N j ) because n j ≤ N j n_j\le\sqrt{N_j} n j ≤ N j . Choose j 0 j_0 j 0 large enough that the complete accepting run, including input access and simulation overhead, satisfies
A M ( x ) ≤ N j when ∣ x ∣ = n j , j ≥ j 0 . (35) A_M(x)\le N_j\qquad\text{when }|x|=n_j,\ j\ge j_0. \tag*{(35)} A M ( x ) ≤ N j when ∣ x ∣ = n j , j ≥ j 0 . (35)
This choice depends on the machine implementation, not on α \alpha α . There is no circularity. The stage generator and the comparisons do not depend on j 0 j_0 j 0 . The final test i ≥ j 0 i\ge j_0 i ≥ j 0 uses a counter in the finite control that saturates at j 0 j_0 j 0 and costs no additional transitions. Hence the constants C ′ , D ′ C',D' C ′ , D ′ do not depend on j 0 j_0 j 0 , and j 0 j_0 j 0 may be chosen after them.
The accepted lengths have arbitrarily large gaps. In fact T j + 1 ≥ N j T_{j+1}\ge\sqrt{N_j} T j + 1 ≥ N j eventually by (A9), and N j + 1 > 2 T j + 1 j + 1 N_{j+1}>2T_{j+1}^{j+1} N j + 1 > 2 T j + 1 j + 1 , so N j + 1 / N j > 2 N j ( j − 1 ) / 2 ⟶ ∞ N_{j+1}/N_j>2N_j^{(j-1)/2}\longrightarrow\infty N j + 1 / N j > 2 N j ( j − 1 ) /2 ⟶ ∞ . Enlarge j 0 j_0 j 0 so that the n j n_j n j are strictly increasing with gaps. The language L G L_G L G is infinite because Σ \Sigma Σ is nonempty, but it contains no cylinder v Σ ∗ v\Sigma^* v Σ ∗ : such a cylinder contains words of every length at least ∣ v ∣ |v| ∣ v ∣ . An infinite reverse-definite language contains a cylinder by (5), since its exceptional set F ′ F' F ′ is finite. Thus L G ∉ R D E F L_G\notin\mathbf{RDEF} L G ∈ / RDEF .
For confidence α \alpha α , take the integer polynomial
Q ( n ) = ⌈ 2 ℓ α ⌉ ( n + 1 ) 4 . Q(n)=\lceil2\ell_\alpha\rceil(n+1)^4. Q ( n ) = ⌈ 2 ℓ α ⌉ ( n + 1 ) 4 .
On an accepted length n = n j n=n_j n = n j , we have N j 3 / 2 < ( n + 1 ) 3 N_j^{3/2}<(n+1)^3 N j 3/2 < ( n + 1 ) 3 , and hence Q ( n ) ≥ 2 ℓ α N j 3 / 2 Q(n)\ge2\ell_\alpha N_j^{3/2} Q ( n ) ≥ 2 ℓ α N j 3/2 . Combining (A6) and (A11) gives A M ( x ) ≤ N j ≤ R α ( Q ( ∣ x ∣ ) ) A_M(x)\le N_j\le R_\alpha(Q(|x|)) A M ( x ) ≤ N j ≤ R α ( Q ( ∣ x ∣ )) . Theorem 2.4 proves L G ∈ P μ , α ∗ L_G\in\mathbf{P}^*_{\mu,\alpha} L G ∈ P μ , α ∗ . The same machine works at every confidence, with only Q Q Q changing, so L G ∈ P μ ∗ L_G\in\mathbf{P}^*_\mu L G ∈ P μ ∗ . The strict upper inclusions follow from L e v e n L_{\mathrm{even}} L even as in the preceding proof. The law is independent of the alphabet, proving the claimed universal quantifier.
Proof of the antichain theorem
For n ≥ 1 n \ge1 n ≥ 1 , let log 2 ( 0 ) n = n \log_2^{(0)} n = n log 2 ( 0 ) n = n and log 2 ( r + 1 ) n = log 2 ( log 2 ( r ) n ) \log_2^{(r+1)} n = \log_2(\log_2^{(r)} n) log 2 ( r + 1 ) n = log 2 ( log 2 ( r ) n ) while the current value exceeds 1 1 1 , and put log ∗ n = min { r ≥ 0 : log 2 ( r ) n ≤ 1 } \log^* n = \min\{r \ge0 : \log_2^{(r)} n \le1\} log ∗ n = min { r ≥ 0 : log 2 ( r ) n ≤ 1 } . Thus log ∗ n \log^* n log ∗ n counts the successive binary logarithms needed to bring n n n to at most 1 1 1 , with log ∗ 1 = 0 \log^* 1 = 0 log ∗ 1 = 0 . The smaller-input case in the proof is handled separately.
Proof. Basic properties. The tower values satisfy tw j ≥ 2 j + 4 \operatorname{tw}_j \ge2^{j+4} tw j ≥ 2 j + 4 , so ∑ j ≥ 1 tw j − 1 ≤ 1 / 16 \sum_{j \ge1} \operatorname{tw}_j^{-1} \le1/16 ∑ j ≥ 1 tw j − 1 ≤ 1/16 . Thus the zero mass is positive and E K ≤ − 1 2 + 1 16 < 0 \mathbb{E}K \le-\frac{1}{2}+\frac{1}{16}<0 E K ≤ − 2 1 + 16 1 < 0 . For s > 0 s>0 s > 0 , E [ ( K + ) s ] = ∑ j ∈ J e tw j 3 s − 4 \mathbb{E}[(K^+)^s]=\sum_{j\in J_e}\operatorname{tw}_j^{3s-4} E [( K + ) s ] = ∑ j ∈ J e tw j 3 s − 4 , which is finite exactly when s < 4 / 3 s<4/3 s < 4/3 . The nonzero point masses are rational, and for k > 0 k>0 k > 0 the value μ ⟨ e ⟩ ( k ) \mu^{\langle e\rangle}(k) μ ⟨ e ⟩ ( k ) is computed by generating tw j \operatorname{tw}_j tw j until tw j 3 ≥ k \operatorname{tw}_j^3\ge k tw j 3 ≥ k , then testing equality and j ∈ J e j\in J_e j ∈ J e . At − 1 -1 − 1 the mass is 1 / 2 1/2 1/2 , and at smaller integers it is zero. The tail of the zero-mass series after stage j j j is at most 2 tw j + 1 − 4 2\operatorname{tw}_{j+1}^{-4} 2 tw j + 1 − 4 . This effective remainder computes the zero mass to any prescribed error and proves computability. A positive atom gives R D E F ⊆ P μ ⟨ e ⟩ , α ∗ \mathbf{RDEF}\subseteq\mathbf{P}^{*}_{\mu^{\langle e\rangle},\alpha} RDEF ⊆ P μ ⟨ e ⟩ , α ∗ by Theorem 2.17, and the finite first positive moment gives containment in P \mathbf{P} P by Theorem 3.1.
The computable zero-drift construction
A slow-stage bound. Fix e e e and an index j ≥ 2 j\ge2 j ≥ 2 with j ∉ J e j\notin J_e j ∈ / J e . We show
Pr μ ⟨ e ⟩ ( τ tw j ≤ tw j ) ≤ ε j a c \Pr_{\mu^{\langle e\rangle}}\left(\tau_{\operatorname{tw}_j}\le\operatorname{tw}_j\right)\le\varepsilon_j^{\mathrm{ac}} μ ⟨ e ⟩ Pr ( τ tw j ≤ tw j ) ≤ ε j ac
: = 6 tw j exp ( − tw j ( log 2 tw j ) 3 ) + 2 tw j 2 − 4 tw j , (36) :=6\operatorname{tw}_j\exp\left(-\frac{\operatorname{tw}_j}{(\log_2\operatorname{tw}_j)^3}\right)+2\operatorname{tw}_j2^{-4\operatorname{tw}_j}, \tag*{(36)} := 6 tw j exp ( − ( log 2 tw j ) 3 tw j ) + 2 tw j 2 − 4 tw j , (36)
and ε j a c → 0 \varepsilon_j^{\mathrm{ac}}\to0 ε j ac → 0 . Put q = tw j q=\operatorname{tw}_j q = tw j . The superscript ac distinguishes the auxiliary quantities here from those in the earlier appendices.
First, the atoms tw i 3 \operatorname{tw}_i^3 tw i 3 with i > j i>j i > j have total mass at most 2 tw j + 1 − 4 = 2 ⋅ 2 − 4 tw j 2\operatorname{tw}_{j+1}^{-4}=2\cdot2^{-4\operatorname{tw}_j} 2 tw j + 1 − 4 = 2 ⋅ 2 − 4 tw j . Replace each such proposal by zero to obtain an iid law ν j a c \nu_j^{\mathrm{ac}} ν j ac . The two held walks agree until the first replaced proposal, so a union bound over the first q q q proposals contributes the second term of (B12).
Second, the positive atoms of ν j a c \nu_j^{\mathrm{ac}} ν j ac are tw i 3 \operatorname{tw}_i^3 tw i 3 with i < j i<j i < j . Since j ∉ J e j\notin J_e j ∈ / J e , all are at most
b j a c = tw j − 1 3 = ( log 2 tw j ) 3 ≥ 32 3 . b_j^{\mathrm{ac}}=\operatorname{tw}_{j-1}^3=(\log_2\operatorname{tw}_j)^3\ge32^3. b j ac = tw j − 1 3 = ( log 2 tw j ) 3 ≥ 3 2 3 .
Put κ j = 1 / b j a c ≤ 1 / 4 \kappa_j=1/b_j^{\mathrm{ac}}\le1/4 κ j = 1/ b j ac ≤ 1/4 . For 0 ≤ x ≤ 1 0\le x\le1 0 ≤ x ≤ 1 , e x ≤ 1 + x + x 2 ≤ 1 + 2 x e^x\le1+x+x^2\le1+2x e x ≤ 1 + x + x 2 ≤ 1 + 2 x , and e − κ j ≤ 1 − κ j + κ j 2 / 2 e^{-\kappa_j}\le1-\kappa_j+\kappa_j^2/2 e − κ j ≤ 1 − κ j + κ j 2 /2 . Hence, using ∑ i < j tw i − 1 ≤ 1 / 16 \sum_{i<j}\operatorname{tw}_i^{-1}\le1/16 ∑ i < j tw i − 1 ≤ 1/16 ,
E ν j a c e κ j K − 1 ≤ − κ j 2 + κ j 2 4 + 2 κ j ∑ i < j tw i − 1 ≤ − κ j 2 + κ j 16 + κ j 8 ≤ − κ j 4 . \begin{aligned}
\mathbb{E}_{\nu_j^{\mathrm{ac}}}e^{\kappa_jK}-1
\le-\frac{\kappa_j}{2}+\frac{\kappa_j^2}{4}+2\kappa_j\sum_{i<j}\operatorname{tw}_i^{-1} \\
&\le-\frac{\kappa_j}{2}+\frac{\kappa_j}{16}+\frac{\kappa_j}{8}
\le-\frac{\kappa_j}{4}.
\end{aligned} E ν j ac e κ j K − 1 ≤ − 2 κ j + 4 κ j 2 + 2 κ j i < j ∑ tw i − 1 ≤ − 2 κ j + 16 κ j + 8 κ j ≤ − 4 κ j .
Write g = E ν j a c e κ j K ≤ 1 − κ j / 4 g=\mathbb{E}_{\nu_j^{\mathrm{ac}}}e^{\kappa_jK}\le1-\kappa_j/4 g = E ν j ac e κ j K ≤ 1 − κ j /4 and V ( i ) = e κ j i V(i)=e^{\kappa_ji} V ( i ) = e κ j i . The only negative proposal is − 1 -1 − 1 , so E [ V ( X t + 1 ) ∣ X t = i ] = g V ( i ) \mathbb{E}[V(X_{t+1})\mid X_t=i]=gV(i) E [ V ( X t + 1 ) ∣ X t = i ] = g V ( i ) for i ≥ 1 i\ge1 i ≥ 1 , and at zero the held proposal adds at most 1 2 ( 1 − e − κ j ) ≤ κ j / 2 \frac12(1-e^{-\kappa_j})\le\kappa_j/2 2 1 ( 1 − e − κ j ) ≤ κ j /2 . Iterating from V ( 0 ) = 1 V(0)=1 V ( 0 ) = 1 gives sup t E V ( X t ) ≤ 1 + ( κ j / 2 ) / ( κ j / 4 ) = 3 \sup_t\mathbb{E}V(X_t)\le1+(\kappa_j/2)/(\kappa_j/4)=3 sup t E V ( X t ) ≤ 1 + ( κ j /2 ) / ( κ j /4 ) = 3 . Markov’s inequality and a union bound give
Pr ν j a c ( τ T ≤ q ) ≤ 3 ( q + 1 ) e − κ j T , \Pr_{\nu_j^{\mathrm{ac}}}(\tau_T\le q)\le3(q+1)e^{-\kappa_jT}, ν j ac Pr ( τ T ≤ q ) ≤ 3 ( q + 1 ) e − κ j T ,
which at T = tw j T=\operatorname{tw}_j T = tw j is at most the first term of (B12), since q ≥ 1 q\ge1 q ≥ 1 .
Finally, j ≤ log 2 tw j j\le\log_2\operatorname{tw}_j j ≤ log 2 tw j , so ln ( tw j ) ≤ ( log 2 tw j ) 2 \ln(\operatorname{tw}_j)\le(\log_2\operatorname{tw}_j)^2 ln ( tw j ) ≤ ( log 2 tw j ) 2 , while tw j / ( log 2 tw j ) 3 \operatorname{tw}_j/(\log_2\operatorname{tw}_j)^3 tw j / ( log 2 tw j ) 3 grows faster than this square. The term 4 ( ln 2 ) tw j 4(\ln2)\operatorname{tw}_j 4 ( ln 2 ) tw j likewise dominates ln ( tw j ) \ln(\operatorname{tw}_j) ln ( tw j ) . Thus both terms of ε j a c \varepsilon_j^{\mathrm{ac}} ε j ac tend to zero.
Membership. A machine M e , j 0 M_{e,j_0} M e , j 0 computes n = ∣ x ∣ n=|x| n = ∣ x ∣ in binary in O ( ( n + 1 ) log ( n + 2 ) ) O((n+1)\log(n+2)) O (( n + 1 ) log ( n + 2 )) transitions, as in Appendix A.1. It then generates tw 1 , tw 2 , … \operatorname{tw}_1,\operatorname{tw}_2,\ldots tw 1 , tw 2 , … in binary in a workspace near the left end. It rejects as soon as some tw j > n \operatorname{tw}_j>n tw j > n , or as soon as tw j < n \operatorname{tw}_j<n tw j < n and tw j \operatorname{tw}_j tw j is at least the bit length of n n n , since then tw j + 1 > n \operatorname{tw}_{j+1}>n tw j + 1 > n . When tw j = n \operatorname{tw}_j=n tw j = n , it accepts exactly if j ∈ J e j\in J_e j ∈ J e and j ≥ j 0 j\ge j_0 j ≥ j 0 , and otherwise rejects. All arithmetic is on numbers of O ( log ( n + 2 ) ) O(\log(n+2)) O ( log ( n + 2 )) bits, and there are at most $\log^* n+1 stages for n ≥ 2 n \ge2 n ≥ 2 ; smaller inputs are handled in constant time. The local arithmetic, including single-tape simulation, has polynomial cost in log ( n + 2 ) \log(n+2) log ( n + 2 ) and is absorbed in the scan bound.
The tests on j j j are implemented in finite control: a residue modulo 2 e + 1 2^{e+1} 2 e + 1 tests j ∈ J e j \in J_e j ∈ J e , and a stage counter saturating at j 0 j_0 j 0 tests j ≥ j 0 j \ge j_0 j ≥ j 0 . Their updates share the existing stage transitions. Taking the product with this finite control adds states but no transitions to the common length-and-tower computation. Hence
A M e , j 0 ( x ) ≤ C ( n + 2 ) log 2 ( n + 2 ) A_{M_{e,j_0}}(x) \le C(n+2)\log_2(n+2) A M e , j 0 ( x ) ≤ C ( n + 2 ) log 2 ( n + 2 )
for a constant C C C independent of both e e e and j 0 j_0 j 0 . Choose j ∗ ≥ 2 j_* \ge2 j ∗ ≥ 2 with C ( t w j + 2 ) log 2 ( t w j + 2 ) ≤ t w j 3 C(\mathrm{tw}_j+2)\log_2(\mathrm{tw}_j+2) \le\mathrm{tw}_j^3 C ( tw j + 2 ) log 2 ( tw j + 2 ) ≤ tw j 3 for all j ≥ j ∗ j \ge j_* j ≥ j ∗ . For ∣ x ∣ = t w j |x|=\mathrm{tw}_j ∣ x ∣ = tw j with j ∈ J e j \in J_e j ∈ J e and j ≥ j 0 ≥ j ∗ j \ge j_0 \ge j_* j ≥ j 0 ≥ j ∗ , a proposal equal to t w j 3 \mathrm{tw}_j^3 tw j 3 crosses the accepting index from every nonnegative position. Hence
Pr ( τ A M e , j 0 ( x ) ≤ q ) ≥ 1 − ( 1 − t w j − 4 ) q ≥ α \Pr(\tau_{A_{M_{e,j_0}}}(x) \le q) \ge1-(1-\mathrm{tw}_j^{-4})^q \ge\alpha Pr ( τ A M e , j 0 ( x ) ≤ q ) ≥ 1 − ( 1 − tw j − 4 ) q ≥ α
whenever q ≥ ℓ α t w j 4 q \ge\ell_\alpha\mathrm{tw}_j^4 q ≥ ℓ α tw j 4 .
The budget Q ( n ) = ⌈ ℓ α ⌉ n 4 Q(n)=\lceil\ell_\alpha\rceil n^4 Q ( n ) = ⌈ ℓ α ⌉ n 4 is therefore adequate, and M e , j 0 M_{e,j_0} M e , j 0 is a decider, so it is sound. Only Q Q Q depends on α \alpha α , so L j 0 ⟨ e ⟩ ∈ P μ ⟨ e ⟩ ∗ L_{j_0}^{\langle e\rangle} \in\mathbf{P}^*_{\mu\langle e\rangle} L j 0 ⟨ e ⟩ ∈ P μ ⟨ e ⟩ ∗ .
Non-membership. Suppose ( M , Q ) (M,Q) ( M , Q ) witnesses L j 0 ⟨ e ⟩ ∈ P μ ⟨ e ′ ⟩ , α ′ ∗ L_{j_0}^{\langle e\rangle} \in\mathbf{P}^*_{\mu\langle e'\rangle,\alpha'} L j 0 ⟨ e ⟩ ∈ P μ ⟨ e ′ ⟩ , α ′ ∗ with e ′ ≠ e e' \ne e e ′ = e . Let j ∈ J e j \in J_e j ∈ J e with j ≥ j 0 j \ge j_0 j ≥ j 0 , choose a ∈ Σ a \in\Sigma a ∈ Σ , and put x = a t w j x=a^{\mathrm{tw}_j} x = a tw j . Then x ∈ L j 0 ⟨ e ⟩ x \in L_{j_0}^{\langle e\rangle} x ∈ L j 0 ⟨ e ⟩ , but x a ∉ L j 0 ⟨ e ⟩ xa \notin L_{j_0}^{\langle e\rangle} x a ∈ / L j 0 ⟨ e ⟩ because t w j + 1 \mathrm{tw}_j+1 tw j + 1 is not a tower value. The distinguishing-cell argument of Appendix A gives A M ( x ) ≥ t w j + 1 A_M(x) \ge\mathrm{tw}_j+1 A M ( x ) ≥ tw j + 1 . For all large j j j , Q ( t w j ) ≤ t w j Q(\mathrm{tw}_j) \le\mathrm{tw}_j Q ( tw j ) ≤ tw j . Since j ∉ J e ′ j \notin J_{e'} j ∈ / J e ′ , (B12) for μ ⟨ e ′ ⟩ \mu\langle e'\rangle μ ⟨ e ′ ⟩ gives
Pr μ ⟨ e ′ ⟩ ( τ A M ( x ) ≤ Q ( t w j ) ) ≤ ε j a c < α ′ \Pr_{\mu\langle e'\rangle}(\tau_{A_M}(x) \le Q(\mathrm{tw}_j)) \le\varepsilon_j^{\mathrm{ac}} < \alpha' μ ⟨ e ′ ⟩ Pr ( τ A M ( x ) ≤ Q ( tw j )) ≤ ε j ac < α ′
for large j ∈ J e j \in J_e j ∈ J e , contradicting completeness.
Conclusion. The language L j 0 ⟨ e ⟩ L_{j_0}^{\langle e\rangle} L j 0 ⟨ e ⟩ is infinite and contains no cylinder v Σ ∗ v\Sigma^* v Σ ∗ , so it is not reverse definite by Theorem 2.16. It belongs to P \mathbf{P} P via M e , j 0 M_{e,j_0} M e , j 0 . Membership and non-membership give pairwise incomparability, also for the intersection classes. The language L j 0 ⟨ e ⟩ L_{j_0}^{\langle e\rangle} L j 0 ⟨ e ⟩ gives strict inclusion above RDEF , and, for any e ′ ≠ e e' \ne e e ′ = e , L j 0 ⟨ e ′ ⟩ ∈ P ∖ P μ ⟨ e ⟩ , α ∗ L_{j_0}^{\langle e'\rangle} \in\mathbf{P} \setminus\mathbf{P}^*_{\mu\langle e\rangle,\alpha} L j 0 ⟨ e ′ ⟩ ∈ P ∖ P μ ⟨ e ⟩ , α ∗ gives strict inclusion below P .
Effectiveness proofs
This appendix proves Theorems 5.9 and 5.11 and records the equivalence between Theorem 2.6 and the standard notion of a computable probability measure on the integers.
On the discrete integers, an effectively open set is a computably enumerable set of integers. The standard characterization of a computable probability measure requires the measures of these sets to be uniformly lower semicomputable: an algorithm produces increasing rational lower bounds converging to the measure [23 ] . Theorem 2.6 implies this property. At stage s ≥ 1 s \ge1 s ≥ 1 , let E s E_s E s be the finite set of distinct integers enumerated so far, indexed by order of first appearance as k j k_j k j . Approximate each μ ( k j ) \mu(k_j) μ ( k j ) in E s E_s E s by d s , j d_{s,j} d s , j with error at most e s , j = 2 − s − j − 1 e_{s,j}=2^{-s-j-1} e s , j = 2 − s − j − 1 , and put L s = ∑ k j ∈ E s max { 0 , d s , j − e s , j } L_s=\sum_{k_j\in E_s}\max\{0,d_{s,j}-e_{s,j}\} L s = ∑ k j ∈ E s max { 0 , d s , j − e s , j } . Then 0 ≤ μ ( E s ) − L s ≤ 2 − s 0\le\mu(E_s)-L_s\le2^{-s} 0 ≤ μ ( E s ) − L s ≤ 2 − s : each clamped summand is at most μ ( k j ) \mu(k_j) μ ( k j ) , and since max { 0 , d s , j − e s , j } ≥ d s , j − e s , j ≥ μ ( k j ) − 2 e s , j \max\{0,d_{s,j}-e_{s,j}\}\ge d_{s,j}-e_{s,j}\ge\mu(k_j)-2e_{s,j} max { 0 , d s , j − e s , j } ≥ d s , j − e s , j ≥ μ ( k j ) − 2 e s , j , the total deficit is at most ∑ j 2 e s , j = 2 − s \sum_j2e_{s,j}=2^{-s} ∑ j 2 e s , j = 2 − s . Since E s ↑ E E_s \uparrow E E s ↑ E , the successive maxima of L s L_s L s increase to μ ( E ) \mu(E) μ ( E ) , uniformly in the enumerator. Conversely, singletons and their complements are effectively open, uniformly in the integer. Lower approximations to both measures, whose sum is one, give upper and lower approximations to each point probability. Searching until their difference is less than 2 − p 2^{-p} 2 − p terminates and gives the presentation in Theorem 2.6. Thus the two definitions agree here.
Proof of Theorem 5.9. Fix a rational a a a with 0 < a < α 0<a<\alpha 0 < a < α . We use a rational a < α a<\alpha a < α because α \alpha α itself need not be computable, and the decider must compare against a fixed rational. We first give an algorithm which, on a positive integer q q q , finds an integer b t a i l ( q ) ≥ 1 b_{\mathrm{tail}}(q)\ge1 b tail ( q ) ≥ 1 with
Pr ( K > b t a i l ( q ) ) < a 2 q . (37) \Pr(K>b_{\mathrm{tail}}(q))<\frac{a}{2q}. \tag*{(37)} Pr ( K > b tail ( q )) < 2 q a . (37)
For s = 1 , 2 , … s=1,2,\ldots s = 1 , 2 , … , approximate every mass in [ − s , s ] ∩ Z [-s,s]\cap\mathbb{Z} [ − s , s ] ∩ Z with error at most 2 − s / ( 2 s + 1 ) 2^{-s}/(2s+1) 2 − s / ( 2 s + 1 ) . The algorithm in Theorem 2.6 provides this by choosing a sufficiently large integer precision. If d k , s d_{k,s} d k , s is the resulting approximation, compute the rational lower bound
L s = ∑ k = − s s max { 0 , d k , s − 2 − s 2 s + 1 } . L_s=\sum_{k=-s}^{s}\max\left\{0,d_{k,s}-\frac{2^{-s}}{2s+1}\right\}. L s = k = − s ∑ s max { 0 , d k , s − 2 s + 1 2 − s } .
Then 0 ≤ μ ( [ − s , s ] ) − L s ≤ 2 1 − s 0\le\mu([-s,s])-L_s\le2^{1-s} 0 ≤ μ ([ − s , s ]) − L s ≤ 2 1 − s . Continuity from below gives μ ( [ − s , s ] ) → 1 \mu([-s,s])\to1 μ ([ − s , s ]) → 1 , so also L s → 1 L_s\to1 L s → 1 . Stop at the first s s s for which the exact rational test L s > 1 − a / ( 2 q ) L_s>1-a/(2q) L s > 1 − a / ( 2 q ) succeeds and put b t a i l ( q ) = s b_{\mathrm{tail}}(q)=s b tail ( q ) = s . The search terminates, and the mass outside [ − s , s ] [-s,s] [ − s , s ] is then less than a / ( 2 q ) a/(2q) a / ( 2 q ) , proving (C13).
If each of the first q q q proposals is at most b t a i l ( q ) b_{\mathrm{tail}}(q) b tail ( q ) , then (4) bounds the maximal position by q b t a i l ( q ) qb_{\mathrm{tail}}(q) q b tail ( q ) . For every integer T > q b t a i l ( q ) T>qb_{\mathrm{tail}}(q) T > q b tail ( q ) , a union bound gives
Pr ( τ T ≤ q ) ≤ q Pr ( K > b t a i l ( q ) ) < a / 2 < α . \Pr(\tau_T\le q)\le q\Pr(K>b_{\mathrm{tail}}(q))<a/2<\alpha. Pr ( τ T ≤ q ) ≤ q Pr ( K > b tail ( q )) < a /2 < α .
Now fix a language witness ( M , Q ) (M,Q) ( M , Q ) at confidence α \alpha α and define the computable integer function
H ( n ) = { 0 , Q ( n ) = 0 , Q ( n ) b t a i l ( Q ( n ) ) , Q ( n ) ≥ 1. H(n)=
\begin{cases}
0, & Q(n)=0,\\
Q(n)b_{\mathrm{tail}}(Q(n)), & Q(n)\ge1.
\end{cases} H ( n ) = { 0 , Q ( n ) b tail ( Q ( n )) , Q ( n ) = 0 , Q ( n ) ≥ 1.
Completeness forces A M ( x ) ≤ H ( ∣ x ∣ ) A_M(x)\le H(|x|) A M ( x ) ≤ H ( ∣ x ∣ ) on every yes-instance. A deterministic decider reads the input to determine its length, computes H ( ∣ x ∣ ) H(|x|) H ( ∣ x ∣ ) , and simulates M M M from its original initial configuration through that logical index, checking index zero as well. It accepts exactly when this simulation accepts and otherwise rejects. The simulation includes every yes-instance’s accepting index, while soundness rules out acceptance on a no-instance. The computation of H H H and the timed simulation both terminate. □
Proof of Theorem 5.11. For every law, soundness and completeness imply that a witnessing machine accepts exactly its language, so the class is contained in RE . For the reverse inclusion, enumerate all machines as M 1 , M 2 , … M_1,M_2,\ldots M 1 , M 2 , … . Let e j e_j e j be the maximum halting index among halting computations of the first j j j machines on inputs of length at most j j j , or zero if none halt. The set is finite. Set E 0 = 0 E_0=0 E 0 = 0 and E j = 1 + max { E j − 1 , e j } E_j=1+\max\{E_{j-1},e_j\} E j = 1 + max { E j − 1 , e j } . This is a halting-time envelope of the classical busy-beaver type [24 Section VI] . Define the law by μ ( E j ) = 1 / ( j ( j + 1 ) ) \mu(E_j)=1/(j(j+1)) μ ( E j ) = 1/ ( j ( j + 1 )) for j ≥ 1 j\ge1 j ≥ 1 , with zero mass elsewhere. The atoms are strictly increasing, the masses sum to one, and Pr ( K ≥ E j ) = 1 / j \Pr(K\ge E_j)=1/j Pr ( K ≥ E j ) = 1/ j .
Fix a recognizing machine M i M_i M i , a confidence α \alpha α , and an accepted input of length n n n . Put j = max { i , n , 1 } ≤ i + n j=\max\{i,n,1\}\le i+n j = max { i , n , 1 } ≤ i + n . Its accepting index is at most e j < E j e_j<E_j e j < E j . Any proposal at least E j E_j E j therefore reaches its accepting tail. Choose an integer k α ≥ max { 1 , ℓ α } k_\alpha\ge\max\{1,\ell_\alpha\} k α ≥ max { 1 , ℓ α } and the integer polynomial Q ( n ) = k α ( i + n ) Q(n)=k_\alpha(i+n) Q ( n ) = k α ( i + n ) . Independent proposals give failure probability at most ( 1 − 1 / j ) Q ( n ) ≤ e − Q ( n ) / j ≤ e − k α ≤ 1 − α (1-1/j)^{Q(n)}\le e^{-Q(n)/j}\le e^{-k_\alpha}\le1-\alpha ( 1 − 1/ j ) Q ( n ) ≤ e − Q ( n ) / j ≤ e − k α ≤ 1 − α . No-instance paths have no accepting index, so the same law realizes every semidecidable language at every fixed confidence. Since such languages include undecidable ones even on a unary alphabet, Theorem 5.9 implies that this law is noncomputable. □
Proof of Theorem 5.10. Theorem 5.9 proves both upper inclusions. For the lower inclusions let M M M be a total decider for L ∈ R E C L\in\mathbf{REC} L ∈ REC . For j ≥ 1 j\ge1 j ≥ 1 , compute the maximum t ( j ) t(j) t ( j ) of T M ( x ) T_M(x) T M ( x ) over the finite set ∣ x ∣ ≤ j |x|\le j ∣ x ∣ ≤ j by running M M M on all those words. Set
E 0 = 0 , E j = 1 + max { E j − 1 , t ( j ) } , μ ( E j ) = 1 j ( j + 1 ) ( j ≥ 1 ) , E_0=0,\qquad E_j=1+\max\{E_{j-1},t(j)\},\qquad\mu(E_j)=\frac{1}{j(j+1)}\quad(j\ge1), E 0 = 0 , E j = 1 + max { E j − 1 , t ( j )} , μ ( E j ) = j ( j + 1 ) 1 ( j ≥ 1 ) ,
with zero mass elsewhere. The atoms are strictly increasing computable positive integers. The masses telescope to one and Pr ( K ≥ ε j ) = 1 / j \Pr(K \geq\varepsilon_j) = 1/j Pr ( K ≥ ε j ) = 1/ j . Given a positive integer k k k , generate the atoms until ε j ≥ k \varepsilon_j \geq k ε j ≥ k , which terminates since ε j ≥ j \varepsilon_j \geq j ε j ≥ j . Return 1 / [ j ( j + 1 ) ] 1/[j(j+1)] 1/ [ j ( j + 1 )] if ε j = k \varepsilon_j = k ε j = k and zero otherwise; nonpositive integers have zero mass. This computes the point probabilities.
On an accepted input of length n n n , put j = max ( 1 , n ) ≤ n + 1 j = \max(1,n) \leq n+1 j = max ( 1 , n ) ≤ n + 1 . Then A M ( x ) ≤ t ( j ) < ε j A_M(x) \leq t(j) < \varepsilon_j A M ( x ) ≤ t ( j ) < ε j . Any proposal at least ε j \varepsilon_j ε j reaches the accepting tail from any nonnegative position. Choose an integer k α ≥ max ( 1 , ℓ α ) k_\alpha\geq\max(1,\ell_\alpha) k α ≥ max ( 1 , ℓ α ) and the polynomial budget Q ( n ) = k α ( n + 1 ) Q(n) = k_\alpha(n+1) Q ( n ) = k α ( n + 1 ) . Its failure probability is at most
( 1 − 1 / j ) Q ( n ) ≤ e − Q ( n ) / j ≤ e − k α ≤ 1 − α . (1 - 1/j)^{Q(n)} \leq e^{-Q(n)/j} \leq e^{-k_\alpha} \leq1-\alpha. ( 1 − 1/ j ) Q ( n ) ≤ e − Q ( n ) / j ≤ e − k α ≤ 1 − α .
Soundness holds because M M M is a decider. The same computable law works at all confidences, with only the budget changing.
Infinite positive moments are not sufficient
The moment obstruction in Theorem 3.1 has no converse for arbitrary laws. We construct a law with all positive moments infinite and class P \mathbf{P} P . Its computability remains open.
Proposition D.1 For every fixed finite nonempty input alphabet Σ \Sigma Σ , there is a law μ \mu μ supported on the positive integers such that E K r = ∞ \mathbb{E}K^r = \infty E K r = ∞ for every r > 0 r > 0 r > 0 , but P μ , α ∗ = P μ ∗ = P \mathbf{P}^{*}_{\mu,\alpha} = \mathbf{P}^{*}_{\mu} = \mathbf{P} P μ , α ∗ = P μ ∗ = P for every $0 < \alpha < 1.
Proof. For a machine M M M , let L ( M ) L(M) L ( M ) be the set of words it eventually accepts. Take a sequence ( M j , Q j , m j ) j ≥ 1 (M_j,Q_j,m_j)_{j\geq1} ( M j , Q j , m j ) j ≥ 1 containing every triple in which L ( M j ) ∉ P L(M_j) \notin\mathbf{P} L ( M j ) ∈ / P , Q j Q_j Q j is an integer polynomial budget, and m j ≥ 2 m_j \geq2 m j ≥ 2 is an integer. Countability gives such a sequence. Start with W 0 = 1 W_0 = 1 W 0 = 1 and ε 0 = 1 / 2 \varepsilon_0 = 1/2 ε 0 = 1/2 . At stage j j j , choose an accepting input x j x_j x j such that, with q j = Q j ( ∣ x j ∣ ) q_j = Q_j(|x_j|) q j = Q j ( ∣ x j ∣ ) ,
A M j ( x j ) > W j − 1 q j . A_{M_j}(x_j) > W_{j-1}q_j. A M j ( x j ) > W j − 1 q j .
Such an input exists. Otherwise simulation through W j − 1 Q j ( n ) W_{j-1}Q_j(n) W j − 1 Q j ( n ) , accepting exactly when M j M_j M j accepts within that bound, would decide L ( M j ) L(M_j) L ( M j ) in polynomial time. Choose a positive rational
ε j ≤ min { ε j − 1 2 , 1 2 m j ( q j + 1 ) } , p j = ε j / 2 , \varepsilon_j \leq\min\left\{\frac{\varepsilon_{j-1}}{2},\frac{1}{2m_j(q_j+1)}\right\}, \qquad p_j = \varepsilon_j/2, ε j ≤ min { 2 ε j − 1 , 2 m j ( q j + 1 ) 1 } , p j = ε j /2 ,
and an integer N j > W j − 1 N_j > W_{j-1} N j > W j − 1 so large that p j N j 1 / j ≥ 1 p_jN_j^{1/j} \geq1 p j N j 1/ j ≥ 1 . Set W j = N j W_j = N_j W j = N j . After completing these choices, define one fixed law by
μ ( N j ) = p j ( j ≥ 1 ) , μ ( 1 ) = 1 − ∑ j ≥ 1 p j . \mu(N_j) = p_j\ (j \geq1), \qquad\mu(1) = 1-\sum_{j\geq1}p_j. μ ( N j ) = p j ( j ≥ 1 ) , μ ( 1 ) = 1 − j ≥ 1 ∑ p j .
All other masses are zero. Since ε j ≤ 2 − j − 1 \varepsilon_j \leq2^{-j-1} ε j ≤ 2 − j − 1 , the sum of the p j p_j p j is at most 1 / 4 1/4 1/4 , so this is a probability law with positive mass at one. Moreover ε ℓ ≤ 2 − ( ℓ − j ) ε j \varepsilon_\ell\leq2^{-(\ell-j)}\varepsilon_j ε ℓ ≤ 2 − ( ℓ − j ) ε j for ℓ ≥ j \ell\geq j ℓ ≥ j , so ∑ ℓ ≥ j p ℓ ≤ ε j \sum_{\ell\geq j}p_\ell\leq\varepsilon_j ∑ ℓ ≥ j p ℓ ≤ ε j . This bound includes all future stages and protects every earlier allowance.
At stage j j j , every support point other than the later points N ℓ N_\ell N ℓ , ℓ ≥ j \ell\geq j ℓ ≥ j , is at most W j − 1 W_{j-1} W j − 1 . If none of the first q j q_j q j proposals uses a later point, their sum is less than A M j ( x j ) A_{M_j}(x_j) A M j ( x j ) . Thus, including when q j = 0 q_j = 0 q j = 0 ,
Pr ( τ A M j ( x j ) ≤ q j ) ≤ q j ∑ ℓ ≥ j p ℓ ≤ q j ε j < 1 / m j . \Pr(\tau_{A_{M_j}}(x_j) \leq q_j) \leq q_j\sum_{\ell\geq j}p_\ell\leq q_j\varepsilon_j < 1/m_j. Pr ( τ A M j ( x j ) ≤ q j ) ≤ q j ℓ ≥ j ∑ p ℓ ≤ q j ε j < 1/ m j .
Any language witness has L = L ( M ) L = L(M) L = L ( M ) , by soundness and completeness. If it witnessed a language outside P \mathbf{P} P at confidence α > 0 \alpha> 0 α > 0 , choose m ≥ 2 m \geq2 m ≥ 2 with 1 / m ≤ α 1/m \leq\alpha 1/ m ≤ α . Its triple ( M , Q , m ) (M,Q,m) ( M , Q , m ) appears in the sequence and the preceding inequality contradicts completeness. Every fixed-confidence class is therefore contained in P \mathbf{P} P . All increments are at least one, so a polynomial-time decider reaches its halting index within its ordinary time bound with probability one. This proves the reverse inclusion at every confidence.
Finally, for any r > 0 r > 0 r > 0 and every integer j ≥ 1 / r j \ge1/r j ≥ 1/ r , p j N j r ≥ p j N j 1 / j ≥ 1 p_jN_j^r \ge p_jN_j^{1/j} \ge1 p j N j r ≥ p j N j 1/ j ≥ 1 . Infinitely many terms of the moment sum are at least one, so E K r = ∞ \mathbb{E}K^r = \infty E K r = ∞ . The decidability bound for computable laws in Theorem 5.9 does not settle the computability of this law: its language class is already contained in P ⊂ \subset ⊂ REC.
Calibrating examples and confidence
These examples calibrate the estimates and isolate the role of confidence and access costs. Each uses the conventions of § 2 .
The moment-bound constant
A one-tick example attains the universal moment-bound coefficient.
Remark E.1 (The numerical constant in the moment bound) For any integer T ≥ 1 T \ge1 T ≥ 1 and 0 < p ≤ 1 0 < p \le1 0 < p ≤ 1 , the law with masses p p p at T T T and 1 − p 1-p 1 − p at zero has m r = p T r m_r = pT^r m r = p T r and Pr ( τ T ≤ q ) = 1 − ( 1 − p ) q \Pr(\tau_T \le q) = 1 - (1-p)^q Pr ( τ T ≤ q ) = 1 − ( 1 − p ) q . At q = 1 q = 1 q = 1 it attains equality in (6) , so the universal coefficient 1 cannot be decreased. This is a family of laws as T T T varies, not an asymptotic equality for one fixed finite-moment law.
Zero-drift alternatives
The fair walk calibrates the quadratic scale and distinguishes the two zero-drift hypotheses.
Example E.2 (The quadratic scale and the two alternatives) For the fair law μ ( − 1 ) = μ ( 1 ) = 1 / 2 \mu(-1) = \mu(1) = 1/2 μ ( − 1 ) = μ ( 1 ) = 1/2 ,
E i τ T = T ( T + 1 ) − i ( i + 1 ) ( 0 ≤ i < T ) , Pr 0 ( τ T ≤ q ) ≤ min { 1 , q / T 2 } ( T ≥ 1 ) . \mathbb{E}_i\tau_T = T(T+1) - i(i+1) \quad(0 \le i < T), \qquad\Pr_0(\tau_T \le q) \le\min\{1,q/T^2\} \quad(T \ge1). E i τ T = T ( T + 1 ) − i ( i + 1 ) ( 0 ≤ i < T ) , 0 Pr ( τ T ≤ q ) ≤ min { 1 , q / T 2 } ( T ≥ 1 ) .
Indeed supported forward blocks give a finite mean hitting time. Writing e i = E i τ T e_i = \mathbb{E}_i\tau_T e i = E i τ T , its first-step recurrence is e T = 0 e_T = 0 e T = 0 , e 0 = 1 + ( e 0 + e 1 ) / 2 e_0 = 1 + (e_0 + e_1)/2 e 0 = 1 + ( e 0 + e 1 ) /2 and e i = 1 + ( e i − 1 + e i + 1 ) / 2 e_i = 1 + (e_{i-1} + e_{i+1})/2 e i = 1 + ( e i − 1 + e i + 1 ) /2 for 0 < i < T 0 < i < T 0 < i < T ; the displayed expression is its unique solution. For the probability bound, the conditional gain of X t 2 X_t^2 X t 2 is 1 at positive states and 1/2 at zero. Summing up to q ∧ τ T q \wedge\tau_T q ∧ τ T gives E 0 X q ∧ τ T 2 ≤ q \mathbb{E}_0X^2_{q\wedge\tau_T} \le q E 0 X q ∧ τ T 2 ≤ q , whereas a hit contributes T 2 T^2 T 2 . Thus α T 2 ≤ q α ( T ) ≤ ⌈ T ( T + 1 ) / ( 1 − α ) ⌉ \alpha T^2 \le q_\alpha(T) \le\lceil T(T+1)/(1-\alpha)\rceil α T 2 ≤ q α ( T ) ≤ ⌈ T ( T + 1 ) / ( 1 − α )⌉ , so q α ( T ) = Θ α ( T 2 ) q_\alpha(T) = \Theta_\alpha(T^2) q α ( T ) = Θ α ( T 2 ) . Conditional blocks sharpen the upper dependence to O ( T 2 ( 1 + ℓ α ) ) O(T^2(1+\ell_\alpha)) O ( T 2 ( 1 + ℓ α )) .
The two sufficient alternatives in Theorem 3.6 are distinct. Let Pr ( W = j ) = j − 2 − ( j + 1 ) − 2 \Pr(W = j) = j^{-2} - (j+1)^{-2} Pr ( W = j ) = j − 2 − ( j + 1 ) − 2 for j ≥ 1 j \ge1 j ≥ 1 and m = E W = ζ ( 2 ) m = \mathbb{E}W = \zeta(2) m = E W = ζ ( 2 ) . Telescoping gives Pr ( W ≥ j ) = j − 2 \Pr(W \ge j) = j^{-2} Pr ( W ≥ j ) = j − 2 ; the tail-sum formula (Fubini–Tonelli theorem [18 Theorem 1.7.2] ) gives finite mean and infinite second moment. Taking K = 1 K = 1 K = 1 with probability m / ( 1 + m ) m/(1+m) m / ( 1 + m ) and K = − W K = -W K = − W otherwise gives zero mean, bounded positive jumps, and infinite variance. In contrast, masses 2 − j − 1 2^{-j-1} 2 − j − 1 at each of j j j and − j -j − j , j ≥ 1 j \ge1 j ≥ 1 , give zero mean and finite variance with unbounded positive jumps. Both laws are computable and have class P by the theorem. For the first, the series for m m m has remainder at most 1 / N 1/N 1/ N after N N N terms, which supplies certified approximations.
Biased nearest-neighbour walk
A nearest-neighbour law calibrates the exponential crossing scale.
Example E.3 (Exponential scale for a biased nearest-neighbor law) Let μ ( 1 ) = p \mu(1) = p μ ( 1 ) = p , μ ( − 1 ) = 1 − p \mu(-1) = 1-p μ ( − 1 ) = 1 − p , where 0 < p < 1 / 2 0 < p < 1/2 0 < p < 1/2 , and put ρ = p / ( 1 − p ) \rho= p/(1-p) ρ = p / ( 1 − p ) . Then q α ( T ) = Θ p , α ( ρ − T ) q_\alpha(T) = \Theta_{p,\alpha}(\rho^{-T}) q α ( T ) = Θ p , α ( ρ − T ) . To see the lower bound, the probabilities π i = ( 1 − ρ ) ρ i \pi_i = (1-\rho)\rho^i π i = ( 1 − ρ ) ρ i are stationary: π i p = π i + 1 ( 1 − p ) \pi_i p = \pi_{i+1}(1-p) π i p = π i + 1 ( 1 − p ) at every edge, including the holding boundary at zero. For these nearest-neighbor updates, i ↦ max ( 0 , i + K ) i \mapsto\max(0,i+K) i ↦ max ( 0 , i + K ) is nondecreasing. Coupling the walk from zero with one started in π \pi π therefore gives Pr 0 ( X t ≥ T ) ≤ ρ T \Pr_0(X_t \ge T) \le\rho^T Pr 0 ( X t ≥ T ) ≤ ρ T , and a union bound gives Pr 0 ( τ T ≤ q ) ≤ q ρ T \Pr_0(\tau_T \le q) \le q\rho^T Pr 0 ( τ T ≤ q ) ≤ q ρ T for T ≥ 1 T \ge1 T ≥ 1 . For the upper bound, supported forward blocks give finite mean. Writing e i = E i τ T e_i = \mathbb{E}_i\tau_T e i = E i τ T , the first-step recurrence gives
E 0 τ T = 1 − p ( 1 − 2 p ) 2 ( ρ − T − 1 ) − T 1 − 2 p . \mathbb{E}_0\tau_T = \frac{1-p}{(1-2p)^2}\left(\rho^{-T}-1\right)-\frac{T}{1-2p}. E 0 τ T = ( 1 − 2 p ) 2 1 − p ( ρ − T − 1 ) − 1 − 2 p T .
For example, writing d i = e i − e i + 1 d_i=e_i-e_{i+1} d i = e i − e i + 1 yields d 0 = 1 / p d_0=1/p d 0 = 1/ p and p d i − ( 1 − p ) d i − 1 = 1 pd_i-(1-p)d_{i-1}=1 p d i − ( 1 − p ) d i − 1 = 1 ; summing the solution up to T − 1 T-1 T − 1 gives this formula. Markov’s inequality completes the quantile upper bound. This calibrates the exponential spatial order in (10) , not its particular choice of κ \kappa κ . The monotone coupling is specific to nearest-neighbor jumps: under the held proposal − 2 -2 − 2 , starts 1 and 2 become 1 and 0.
Beyond exponential moments
The all-moments classification applies beyond exponential moments.
Example E.4 (Beyond exponential moments) Let
μ ( − 1 ) = 1 2 , μ ( 2 j ) = 2 − j 2 − 2 j − 4 ( j ≥ 1 ) , μ ( 0 ) = 1 2 − ∑ j ≥ 1 2 − j 2 − 2 j − 4 . \mu(-1)=\frac{1}{2},\qquad\mu(2^j)=2^{-j^2-2j-4}\ (j\ge1),\qquad\mu(0)=\frac{1}{2}-\sum_{j\ge1}2^{-j^2-2j-4}. μ ( − 1 ) = 2 1 , μ ( 2 j ) = 2 − j 2 − 2 j − 4 ( j ≥ 1 ) , μ ( 0 ) = 2 1 − j ≥ 1 ∑ 2 − j 2 − 2 j − 4 .
The positive mass is less than 1 / 2 1/2 1/2 . Its positive mean is at most ∑ j ≥ 1 2 − 2 j − 4 = 1 / 48 \sum_{j\ge1}2^{-2j-4}=1/48 ∑ j ≥ 1 2 − 2 j − 4 = 1/48 , using j 2 ≥ j j^2\ge j j 2 ≥ j , whereas E K − = 1 / 2 \mathbb{E}K^-=1/2 E K − = 1/2 . Every positive moment is finite because ∑ j ≥ 1 2 − j 2 + ( s − 2 ) j − 4 < ∞ \sum_{j\ge1}2^{-j^2+(s-2)j-4}<\infty ∑ j ≥ 1 2 − j 2 + ( s − 2 ) j − 4 < ∞ for each s > 0 s>0 s > 0 . For any κ > 0 \kappa>0 κ > 0 , the terms e κ 2 j 2 − j 2 − 2 j − 4 e^{\kappa2^j}2^{-j^2-2j-4} e κ 2 j 2 − j 2 − 2 j − 4 do not tend to zero, so E e κ K = ∞ \mathbb{E}e^{\kappa K}=\infty E e κ K = ∞ . The law is computable: all nonzero point masses are rational, and the remainder of the zero-mass series after N N N terms is at most ∑ j > N 2 − 2 j − 4 = 2 − 2 N − 4 / 3 \sum_{j>N}2^{-2j-4}=2^{-2N-4}/3 ∑ j > N 2 − 2 j − 4 = 2 − 2 N − 4 /3 . Theorem 4.1 gives class RDEF, while the exponential-contraction criterion does not apply.
One-jump budgets
The next comparison identifies when a one-jump budget has the correct order.
Remark E.5 (When a one-jump budget has the right order) Fix 0 < a < 1 0<a<1 0 < a < 1 and the positive law μ ( j ) = j − a − ( j + 1 ) − a \mu(j)=j^{-a}-(j+1)^{-a} μ ( j ) = j − a − ( j + 1 ) − a , j ≥ 1 j\ge1 j ≥ 1 . Its exact tail is T − a T^{-a} T − a . If ∑ j = 1 q K j ≥ T \sum_{j=1}^q K_j\ge T ∑ j = 1 q K j ≥ T , then ∑ j = 1 q min ( K j , T ) ≥ T \sum_{j=1}^q\min(K_j,T)\ge T ∑ j = 1 q min ( K j , T ) ≥ T . The tail-sum formula and an integral comparison give E min ( K , T ) = ∑ m = 1 T m − a ≤ T 1 − a / ( 1 − a ) \mathbb{E}\min(K,T)=\sum_{m=1}^T m^{-a}\le T^{1-a}/(1-a) E min ( K , T ) = ∑ m = 1 T m − a ≤ T 1 − a / ( 1 − a ) . Markov’s inequality therefore bounds the hitting probability by q T − a / ( 1 − a ) qT^{-a}/(1-a) q T − a / ( 1 − a ) . The one-jump lower bound from Theorem 4.4 gives q α ( T ) = Θ a , α ( T a ) q_\alpha(T)=\Theta_{a,\alpha}(T^a) q α ( T ) = Θ a , α ( T a ) for this fixed law. For the same tail form with a > 1 a>1 a > 1 , the mean is finite. The first-moment bound gives q α ( T ) ≥ α T / E K q_\alpha(T)\ge\alpha T/\mathbb{E}K q α ( T ) ≥ α T / E K , while K ≥ 1 K\ge1 K ≥ 1 gives q α ( T ) ≤ T q_\alpha(T)\le T q α ( T ) ≤ T . Thus the one-jump T a T^a T a budget can be nonoptimal. Its tail hypothesis is not necessary for class P either: δ 1 \delta_1 δ 1 already gives equality.
Sampling and access cost
Exact sampling and charged access have different costs.
Remark E.6 (Effectiveness and access) The law μ E \mu_E μ E has computable rational point masses and an effective tail. It can be sampled exactly with fair bits in almost surely finite time. For example, take U U U uniform on ( 0 , 1 ) (0,1) ( 0 , 1 ) , determine J = ⌊ 1 / U ⌋ J=\lfloor1/U\rfloor J = ⌊ 1/ U ⌋ by reading successive binary digits until the containing interval ( 1 / ( j + 1 ) , 1 / j ] (1/(j+1),1/j] ( 1/ ( j + 1 ) , 1/ j ] is determined, and output K = 2 J K=2^J K = 2 J . Except at the countably many endpoints, this determination takes finitely many digits; those endpoints have probability zero. The interval lengths give Pr ( J = j ) = 1 / ( j ( j + 1 ) ) \Pr(J=j)=1/(j(j+1)) Pr ( J = j ) = 1/ ( j ( j + 1 )) .
Writing a full binary search for K K K has infinite expected cost: its length is J + 1 J+1 J + 1 , and E J = ∑ m ≥ 1 Pr ( J ≥ m ) = ∑ m ≥ 1 1 / m = ∞ \mathbb{E}J=\sum_{m\ge1}\Pr(J\ge m)=\sum_{m\ge1}1/m=\infty E J = ∑ m ≥ 1 Pr ( J ≥ m ) = ∑ m ≥ 1 1/ m = ∞ . The EXP classification uses the free configuration access of Section 2.
Confidence for position-dependent clocks
Different confidence levels can give different classes for a position-dependent kernel.
Since X q ≥ q X_q \ge q X q ≥ q surely, polynomial-time deciders have a sure polynomial budget. If α < 1 / 2 \alpha< 1/2 α < 1/2 , choose an integer k ≥ max { 1 , ℓ 2 α } k \ge\max\{1,\ell_{2\alpha}\} k ≥ max { 1 , ℓ 2 α } and use Q ( n ) = 1 + k p ( n ) Q(n) = 1+kp(n) Q ( n ) = 1 + k p ( n ) for an exponential logical bound 2 p ( n ) 2^{p(n)} 2 p ( n ) , p ≥ 1 p \ge1 p ≥ 1 . Conditional on the even lane, one of its k p ( n ) kp(n) k p ( n ) samples crosses this bound with probability at least 2 α 2\alpha 2 α , by (13) . Conversely, unused independent samples give max t ≤ q X t ≤ 4 q + 2 + ∑ j ≤ q Y j \max_{t \le q} X_t \le4q+2+\sum_{j \le q}Y_j max t ≤ q X t ≤ 4 q + 2 + ∑ j ≤ q Y j . The union-bound argument of Theorem 6.10 forces T ≤ 4 q + 2 + q 2 q / α T \le4q+2+q^{2q/\alpha} T ≤ 4 q + 2 + q 2 q / α at confidence α \alpha α , giving an exponential logical-time ceiling for a polynomial budget. For finite q ≥ 1 q \ge1 q ≥ 1 and T > 4 q + 2 T > 4q+2 T > 4 q + 2 , the slow lane fails, and the fast lane has positive failure probability: all its q − 1 q-1 q − 1 samples can equal two. Total success is therefore strictly below 1 / 2 1/2 1/2 at every such finite horizon. At confidence at least 1 / 2 1/2 1/2 , accepting time is therefore at most 4 Q ( n ) + 2 4Q(n)+2 4 Q ( n ) + 2 , giving a polynomial-time simulation.
Confidence transfer for iid clocks
For iid laws, a restart comparison transfers several radius bounds between confidence levels.
Common budgets over parameter sets
We ask whether one machine and one budget work for every fixed clock in a nonempty allowed parameter set. For the two families below, the criterion depends on the set’s infimum.
Let Θ \Theta Θ denote the parameter domain of the clock family, and let its possibility distribution π : Θ → [ 0 , 1 ] \pi:\Theta\to[0,1] π : Θ → [ 0 , 1 ] assign a degree of possibility to each fixed parameter. For a cut level γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) , write Θ γ = { θ ∈ Θ : π ( θ ) ≥ γ } \Theta_{\gamma} = \{\theta\in\Theta: \pi(\theta) \ge\gamma\} Θ γ = { θ ∈ Θ : π ( θ ) ≥ γ } ; we consider nonempty cuts [6 ] . These grades specify allowed fixed parameter values; the probability law governs the clock at the chosen parameter throughout a run, rather than choosing a new parameter at every tick. The lazy-speed domain is Θ = [ 0 , 1 ] \Theta= [0,1] Θ = [ 0 , 1 ] , with positive-speed examples restricted to ( 0 , 1 ] (0,1] ( 0 , 1 ] ; the activation domain is Θ = [ 0 , 1 / 2 ] \Theta= [0,1/2] Θ = [ 0 , 1/2 ] , with the positive-return example restricted to ( 0 , 1 / 2 ] (0,1/2] ( 0 , 1/2 ] . Only the infimum matters here: π ( p ) = p \pi(p) = p π ( p ) = p gives lazy cuts [ γ , 1 ] [\gamma,1] [ γ , 1 ] with linear budgets, while a constant possibility on ( 0 , 1 ] (0,1] ( 0 , 1 ] allows speeds approaching zero and no common finite budget. For parity clocks, a cut ( 0 , 1 / 2 ] (0,1/2] ( 0 , 1/2 ] has common class P although each member has class EXP , whereas π ( β ) = 2 β \pi(\beta) = 2\beta π ( β ) = 2 β gives positive-infimum cuts [ γ / 2 , 1 / 2 ] [\gamma/2,1/2] [ γ /2 , 1/2 ] with common class EXP .
A lazy-clock benchmark
At every fixed positive speed, the lazy clock has class P . A family of such clocks can fail to admit a common finite budget. The family is
μ p = ( 1 − p ) δ 0 + p δ 1 , 0 ≤ p ≤ 1 , (38) \mu_p = (1-p)\delta_0 + p\delta_1,\qquad0 \le p \le1, \tag*{(38)} μ p = ( 1 − p ) δ 0 + p δ 1 , 0 ≤ p ≤ 1 , (38)
with the held update (1). Its position after Q Q Q ticks is binomial with parameters Q Q Q and p p p ; there are no backward proposals. Let S ⊆ [ 0 , 1 ] \mathcal{S} \subseteq[0,1] S ⊆ [ 0 , 1 ] be a nonempty allowed set, for instance a cut Θ γ \Theta_\gamma Θ γ , put p ∗ = inf S p_* = \inf\mathcal{S} p ∗ = inf S , and define the common-budget class U S , α \mathcal{U}_{\mathcal{S},\alpha} U S , α by Theorem 2.3 with the budget condition replaced by
inf p ∈ S Pr ( p ) ( X Q ( ∣ x ∣ ) ≥ A M ( x ) ) ≥ α on yes-instances , \inf_{p\in\mathcal{S}} \Pr^{(p)}(X_Q(|x|) \ge A_M(x)) \ge\alpha\quad\text{on yes-instances}, p ∈ S inf Pr ( p ) ( X Q ( ∣ x ∣ ) ≥ A M ( x )) ≥ α on yes-instances ,
one machine and one budget for all p ∈ S p \in\mathcal{S} p ∈ S . Also define q α , S ( T ) = min { Q ∈ N : inf p ∈ S Pr ( p ) ( X Q ≥ T ) ≥ α } q_{\alpha,\mathcal{S}}(T) = \min\{Q \in\mathbb{N} : \inf_{p\in\mathcal{S}} \Pr^{(p)}(X_Q \ge T) \ge\alpha\} q α , S ( T ) = min { Q ∈ N : inf p ∈ S Pr ( p ) ( X Q ≥ T ) ≥ α } . Because this walk never decreases, the event is threshold crossing by tick Q Q Q . The criterion depends on p ∗ p_* p ∗ .
Proof sketch. Couple successes with common uniforms; binomial crossing probabilities are nondecreasing and continuous in p p p , so their infimum is the value at p ∗ p_* p ∗ , attained or not. For p ∗ > 0 p_* > 0 p ∗ > 0 , Markov’s inequality gives both displayed bounds, and H T H_T H T is a sum of T T T geometric waits with mean T / p T/p T / p and variance T ( 1 − p ) / p 2 T(1-p)/p^2 T ( 1 − p ) / p 2 . Chebyshev’s bound Pr ( ∣ H T − T / p ∣ ≥ ε T / p ) ≤ 1 / ( ε 2 T ) \Pr(|H_T-T/p|\ge\varepsilon T/p)\le1/(\varepsilon^2T) Pr ( ∣ H T − T / p ∣ ≥ εT / p ) ≤ 1/ ( ε 2 T ) is uniform in p > 0 p>0 p > 0 ; squeezing the quantiles and allowing rounding error p ∗ / T ≤ 1 / T p_*/T\le1/T p ∗ / T ≤ 1/ T proves the uniform limit. When p ∗ = 0 p_* = 0 p ∗ = 0 , Pr ( p ) ( X Q ≥ T ) ≤ Q p \Pr^{(p)}(X_Q\ge T)\le Qp Pr ( p ) ( X Q ≥ T ) ≤ Qp for T ≥ 1 T\ge1 T ≥ 1 , so no common finite budget exists. Bounded jumps give the upper class bound by Theorem 6.5, and a positive infimum supplies a common polynomial budget to any polynomial-time decider; at zero infimum, Theorem 2.12’s time-zero argument gives only the trivial languages.
Common budgets for the activation family
For the activation clocks of Section 6.2, let the allowed backward probabilities form a nonempty set B ⊆ [ 0 , 1 / 2 ] \mathcal{B} \subseteq[0,1/2] B ⊆ [ 0 , 1/2 ] and put b ∗ = inf B b_* = \inf\mathcal{B} b ∗ = inf B . A possibility cut is one such set.
Definition F.2 (Common-budget class for the activation family) For 0 < α < 1 0<\alpha<1 0 < α < 1 , a language L L L belongs to U B , α \mathcal{U}_{\mathcal{B},\alpha} U B , α if some deterministic machine M M M and integer polynomial Q Q Q satisfy A M ( x ) = ∞ A_M(x)=\infty A M ( x ) = ∞ on every x ∉ L x\notin L x ∈ / L , and, on every x ∈ L x\in L x ∈ L , A M ( x ) < ∞ A_M(x)<\infty A M ( x ) < ∞ and
inf β ∈ B Pr ( β ) ( τ A M ( x ) ≤ Q ( ∣ x ∣ ) ) ≥ α . \inf_{\beta\in\mathcal{B}}\Pr^{(\beta)}(\tau_{A_M(x)}\le Q(|x|))\ge\alpha. β ∈ B inf Pr ( β ) ( τ A M ( x ) ≤ Q ( ∣ x ∣ )) ≥ α . One machine and one budget serve every β ∈ B \beta\in\mathcal{B} β ∈ B . Also let q α , B ( T ) = min { q ∈ N : inf β ∈ B Pr ( β ) ( τ T ≤ q ) ≥ α } q_{\alpha,\mathcal{B}}(T)=\min\{q\in\mathbb{N}:\inf_{\beta\in\mathcal{B}}\Pr^{(\beta)}(\tau_T\le q)\ge\alpha\} q α , B ( T ) = min { q ∈ N : inf β ∈ B Pr ( β ) ( τ T ≤ q ) ≥ α } .
Theorem F.3 (Common budgets over backward-probability cuts) Let B ⊆ [ 0 , 1 / 2 ] \mathcal{B}\subseteq[0,1/2] B ⊆ [ 0 , 1/2 ] be nonempty with b ∗ = inf B b_*=\inf\mathcal{B} b ∗ = inf B , and fix 0 < α < 1 0<\alpha<1 0 < α < 1 . Then:
(i) if b ∗ > 0 b_*>0 b ∗ > 0 , then U B , α = E X P \mathcal{U}_{\mathcal{B},\alpha}=\mathbf{EXP} U B , α = EXP and
lim T → ∞ q α , B ( T ) log 2 T = ( 1 + 1 b ∗ ) ℓ α ; \lim_{T\to\infty}\frac{q_{\alpha,\mathcal{B}}(T)}{\log_2 T}=\left(1+\frac{1}{b_*}\right)\ell_\alpha; T → ∞ lim log 2 T q α , B ( T ) = ( 1 + b ∗ 1 ) ℓ α ; (ii) if b ∗ = 0 b_*=0 b ∗ = 0 , then U B , α = P \mathcal{U}_{\mathcal{B},\alpha}=\mathbf{P} U B , α = P and q α , B ( T ) = q α ( 0 ) ( T ) q_{\alpha,\mathcal{B}}(T)=q^{(0)}_\alpha(T) q α , B ( T ) = q α ( 0 ) ( T ) for every T ∈ N T\in\mathbb{N} T ∈ N , and q α , B ( T ) / T → 1 q_{\alpha,\mathcal{B}}(T)/T\to1 q α , B ( T ) / T → 1 .
Constants may depend on B \mathcal{B} B and α \alpha α ; in (ii) they may be chosen to depend only on α \alpha α . Neither the attainment of the infimum nor the boundary convention of the cut is used.
Every fixed β > 0 \beta>0 β > 0 has C η β , α = E X P C_{\eta\beta,\alpha}=\mathbf{EXP} C η β , α = EXP by Theorem 6.6 ; the theorem gives common class P \mathbf{P} P when the allowed positive values have infimum zero. The proof rests on finite-horizon continuity, a comparison of running maxima with the zero-parameter clock, and the resulting expected crossing-time bound uniform in β \beta β .
Lemma F.4 (Finite-horizon parameter continuity) For every β , β ′ ∈ [ 0 , 1 / 2 ] \beta,\beta'\in[0,1/2] β , β ′ ∈ [ 0 , 1/2 ] , q ≥ 0 q\ge0 q ≥ 0 , and T ∈ N T\in\mathbb{N} T ∈ N ,
∣ Pr ( β ) ( τ T ≤ q ) − Pr ( β ′ ) ( τ T ≤ q ) ∣ ≤ min { 1 , q ∣ β − β ′ ∣ } . \left|\Pr^{(\beta)}(\tau_T\le q)-\Pr^{(\beta')}(\tau_T\le q)\right|\le\min\{1,q|\beta-\beta'|\}. Pr ( β ) ( τ T ≤ q ) − Pr ( β ′ ) ( τ T ≤ q ) ≤ min { 1 , q ∣ β − β ′ ∣ } . Consequently, for nonempty B \mathcal{B} B and b ∗ = inf B b_*=\inf\mathcal{B} b ∗ = inf B , inf β ∈ B Pr ( β ) ( τ T ≤ q ) ≤ Pr ( b ∗ ) ( τ T ≤ q ) \inf_{\beta\in\mathcal{B}}\Pr^{(\beta)}(\tau_T\le q)\le\Pr^{(b_*)}(\tau_T\le q) inf β ∈ B Pr ( β ) ( τ T ≤ q ) ≤ Pr ( b ∗ ) ( τ T ≤ q ) .
Proof. Rows at zero and at odd states agree. At each positive even state the rows transfer mass ∣ β − β ′ ∣ |\beta-\beta'| ∣ β − β ′ ∣ between the hold and the move − 1 -1 − 1 ; their total variation distance is ∣ β − β ′ ∣ |\beta-\beta'| ∣ β − β ′ ∣ . Whenever the histories agree, couple the next states maximally. A union bound bounds the chance of disagreement by tick q q q by q ∣ β − β ′ ∣ q|\beta-\beta'| q ∣ β − β ′ ∣ . On agreement, the threshold-crossing events agree. Probabilities differ by at most one as well. Since b ∗ b_* b ∗ is in the closure of B \mathcal{B} B , take a sequence of allowed parameters tending to it. Continuity at the fixed horizon gives the upper bound on the infimum. □
Lemma F.5 (Comparison with the zero-parameter clock) For every β ∈ [ 0 , 1 / 2 ] \beta\in[0,1/2] β ∈ [ 0 , 1/2 ] , q , T ∈ N q,T\in\mathbb{N} q , T ∈ N , Pr ( β ) ( τ T ≤ q ) ≥ Pr ( 0 ) ( τ T ≤ q ) \Pr^{(\beta)}(\tau_T\le q)\ge\Pr^{(0)}(\tau_T\le q) Pr ( β ) ( τ T ≤ q ) ≥ Pr ( 0 ) ( τ T ≤ q ) . If inf B = 0 \inf\mathcal{B}=0 inf B = 0 , their lower envelope is exactly inf β ∈ B Pr ( β ) ( τ T ≤ q ) = Pr ( 0 ) ( τ T ≤ q ) \inf_{\beta\in\mathcal{B}}\Pr^{(\beta)}(\tau_T\le q)=\Pr^{(0)}(\tau_T\le q) inf β ∈ B Pr ( β ) ( τ T ≤ q ) = Pr ( 0 ) ( τ T ≤ q ) .
Proof. Write X t ( 0 ) X_t^{(0)} X t ( 0 ) for the baseline. Couple both clocks’ first two ticks using the same initial heavy sample, so both positions at tick two are 2 + Y 2+Y 2 + Y and are even. Subsequently use independent fair coins, one per tick: the baseline moves + 2 +2 + 2 on success and holds on failure. At an even state the β \beta β -clock also moves + 2 +2 + 2 on success; on failure it returns by one with conditional probability 2 β 2\beta 2 β and otherwise holds. This gives the required probabilities 1 / 2 , β , 1 / 2 − β 1/2,\beta,1/2-\beta 1/2 , β , 1/2 − β . At an odd state it takes its mandatory displacement 1 + Y ′ ≥ 3 1+Y'\ge3 1 + Y ′ ≥ 3 with a fresh independent heavy sample. All extra choices are independent of the coins and earlier samples, so each marginal has its stated kernel.
Inductively, at every t ≥ 2 t\ge2 t ≥ 2 an even β \beta β -position is at least X t ( 0 ) X_t^{(0)} X t ( 0 ) , and an odd β \beta β -position is at least X t ( 0 ) − 1 X_t^{(0)}-1 X t ( 0 ) − 1 . From an even state, coin success preserves the comparison; coin failure holds the baseline and either holds the other clock or lowers it by one. From an odd state, the displacement at least three restores the even comparison even if the baseline moves by two. If the β \beta β -position is one below the baseline, its preceding even position was at least that baseline position: the baseline held on the return tick. Thus at every tick its running maximum is at least the baseline position, which is itself nondecreasing. The running maxima, and hence the hitting events, are ordered. The equality of lower envelopes when inf B = 0 \inf\mathcal{B}=0 inf B = 0 follows from this inequality and Theorem F.4.
Lemma F.6 (Uniform linear expected crossing time) From zero, for every β ∈ [ 0 , 1 / 2 ] \beta\in[0,1/2] β ∈ [ 0 , 1/2 ] and integer T ≥ 1 T \ge1 T ≥ 1 ,
E ( β ) τ T ≤ E ( 0 ) τ T ≤ 2 T , Pr ( τ T > q ) ≤ 2 T q ( q ≥ 1 ) . \mathbb{E}^{(\beta)}\tau_T \le\mathbb{E}^{(0)}\tau_T \le2T,\qquad\Pr(\tau_T > q) \le\frac{2T}{q}\quad(q \ge1). E ( β ) τ T ≤ E ( 0 ) τ T ≤ 2 T , Pr ( τ T > q ) ≤ q 2 T ( q ≥ 1 ) .
Proof. Theorem F.5 gives Pr ( β ) ( τ T > q ) ≤ Pr ( 0 ) ( τ T > q ) \Pr^{(\beta)}(\tau_T > q) \le\Pr^{(0)}(\tau_T > q) Pr ( β ) ( τ T > q ) ≤ Pr ( 0 ) ( τ T > q ) at every integer q ≥ 0 q \ge0 q ≥ 0 . Sum these tails. Tonelli’s theorem for nonnegative summands [18 Theorem 1.7.2] gives
E ( β ) τ T = ∑ q = 0 ∞ Pr ( β ) ( τ T > q ) ≤ ∑ q = 0 ∞ Pr ( 0 ) ( τ T > q ) = E ( 0 ) τ T ≤ 2 T , \mathbb{E}^{(\beta)}\tau_T = \sum_{q=0}^{\infty}\Pr^{(\beta)}(\tau_T > q) \le\sum_{q=0}^{\infty}\Pr^{(0)}(\tau_T > q) = \mathbb{E}^{(0)}\tau_T \le2T, E ( β ) τ T = q = 0 ∑ ∞ Pr ( β ) ( τ T > q ) ≤ q = 0 ∑ ∞ Pr ( 0 ) ( τ T > q ) = E ( 0 ) τ T ≤ 2 T ,
where the last inequality is Theorem 6.8. Markov’s inequality proves the tail bound.
Proof of Theorem F.3(ii). First, B \mathcal{B} B contains a sequence β j → b ∗ = 0 \beta_j \to b_* = 0 β j → b ∗ = 0 (or contains zero). Let ( M , Q ) (M,Q) ( M , Q ) witness L ∈ U B , α L \in\mathcal{U}_{\mathcal{B},\alpha} L ∈ U B , α , and fix a yes-instance x x x with budget q = Q ( ∣ x ∣ ) q = Q(|x|) q = Q ( ∣ x ∣ ) . For every j j j , Pr ( β j ) ( τ A M ( x ) ≤ q ) ≥ α \Pr^{(\beta_j)}(\tau_{A_M(x)} \le q) \ge\alpha Pr ( β j ) ( τ A M ( x ) ≤ q ) ≥ α , and Theorem F.4 gives Pr ( 0 ) ( τ A M ( x ) ≤ q ) ≥ α − q β j \Pr^{(0)}(\tau_{A_M(x)} \le q) \ge\alpha-q\beta_j Pr ( 0 ) ( τ A M ( x ) ≤ q ) ≥ α − q β j . Letting j → ∞ j \to\infty j → ∞ shows that the same ( M , Q ) (M,Q) ( M , Q ) witnesses L ∈ C η 0 , α = P L \in\mathcal{C}_{\eta_0,\alpha} = \mathbf{P} L ∈ C η 0 , α = P . Hence U B , α ⊆ P \mathcal{U}_{\mathcal{B},\alpha} \subseteq\mathbf{P} U B , α ⊆ P .
Conversely, let M M M decide L L L within the integer polynomial bound H ( n ) H(n) H ( n ) . Choose an integer k α , l i n ≥ 2 / ( 1 − α ) k_{\alpha,\mathrm{lin}} \ge2/(1-\alpha) k α , lin ≥ 2/ ( 1 − α ) and set Q ( n ) = k α , l i n ( H ( n ) + 1 ) Q(n) = k_{\alpha,\mathrm{lin}}(H(n)+1) Q ( n ) = k α , lin ( H ( n ) + 1 ) , an integer polynomial with Q ( n ) ≥ 1 Q(n) \ge1 Q ( n ) ≥ 1 . On a yes-instance with accepting index T = A M ( x ) ≤ H ( n ) T = A_M(x) \le H(n) T = A M ( x ) ≤ H ( n ) : if T = 0 T=0 T = 0 then τ 0 = 0 \tau_0=0 τ 0 = 0 ; otherwise Theorem F.6 and monotonicity in the target give, for every β ∈ [ 0 , 1 / 2 ] \beta\in[0,1/2] β ∈ [ 0 , 1/2 ] ,
Pr ( β ) ( τ T > Q ( n ) ) ≤ 2 T Q ( n ) ≤ 2 H ( n ) k α , l i n ( H ( n ) + 1 ) ≤ 1 − α . \Pr^{(\beta)}(\tau_T > Q(n)) \le\frac{2T}{Q(n)} \le\frac{2H(n)}{k_{\alpha,\mathrm{lin}}(H(n)+1)} \le1-\alpha. Pr ( β ) ( τ T > Q ( n )) ≤ Q ( n ) 2 T ≤ k α , lin ( H ( n ) + 1 ) 2 H ( n ) ≤ 1 − α .
The same M , Q M,Q M , Q therefore work uniformly over every B \mathcal{B} B , and P ⊆ U B , α \mathbf{P} \subseteq\mathcal{U}_{\mathcal{B},\alpha} P ⊆ U B , α .
By Theorem F.5, the lower envelope equals the baseline hitting probability at every finite horizon, so q α , B ( T ) = q 0 α ( T ) q_{\alpha,\mathcal{B}}(T) = q^{\alpha}_0(T) q α , B ( T ) = q 0 α ( T ) for all T ∈ N T \in\mathbb{N} T ∈ N . The limit in Theorem 6.11 gives the exact linear leading constant.
Proof of Theorem F.3(i). Let b ∗ > 0 b_* > 0 b ∗ > 0 , so every β ∈ B \beta\in\mathcal{B} β ∈ B lies in [ b ∗ , 1 / 2 ] [b_*,1/2] [ b ∗ , 1/2 ] .
For E X P ⊆ U B , α \mathbf{EXP} \subseteq\mathcal{U}_{\mathcal{B},\alpha} EXP ⊆ U B , α , let M M M decide L L L within H ( n ) = 2 p ( n ) H(n)=2^{p(n)} H ( n ) = 2 p ( n ) transitions, where p p p has nonnegative integer coefficients and p ( n ) ≥ 1 p(n) \ge1 p ( n ) ≥ 1 for every n ∈ N n \in\mathbb{N} n ∈ N . Choose an integer k α , B ≥ max { 1 , b ∗ − 1 ℓ α } k_{\alpha,\mathcal{B}} \ge\max\{1,b_*^{-1}\ell_\alpha\} k α , B ≥ max { 1 , b ∗ − 1 ℓ α } and put Q ( n ) = 2 k α , B p ( n ) Q(n)=2k_{\alpha,\mathcal{B}}p(n) Q ( n ) = 2 k α , B p ( n ) , an allowed polynomial. For any β ∈ B \beta\in\mathcal{B} β ∈ B and any integer r r r with 2 r ≤ Q ( n ) 2r \le Q(n) 2 r ≤ Q ( n ) , Theorem 6.9 gives, from every state below a target T ≥ 1 T \ge1 T ≥ 1 ,
Pr ( β ) ( τ T > 2 r ) ≤ exp ( − β r h ( T ) ) ≤ exp ( − b ∗ r h ( T ) ) . \Pr^{(\beta)}(\tau_T > 2r) \le\exp\left(-\frac{\beta r}{h(T)}\right) \le\exp\left(-\frac{b_*r}{h(T)}\right). Pr ( β ) ( τ T > 2 r ) ≤ exp ( − h ( T ) β r ) ≤ exp ( − h ( T ) b ∗ r ) .
Take T = H ( n ) T=H(n) T = H ( n ) , so h ( T ) = p ( n ) h(T)=p(n) h ( T ) = p ( n ) , and r = k α , B p ( n ) ∈ N r=k_{\alpha,\mathcal{B}}p(n)\in\mathbb{N} r = k α , B p ( n ) ∈ N . Then 2 r ≤ Q ( n ) 2r\le Q(n) 2 r ≤ Q ( n ) and b ∗ r ≥ b ∗ k α , B p ( n ) ≥ p ( n ) ℓ α b_*r\ge b_*k_{\alpha,\mathcal{B}}p(n)\ge p(n)\ell_\alpha b ∗ r ≥ b ∗ k α , B p ( n ) ≥ p ( n ) ℓ α , so Pr ( β ) ( τ H ( n ) > Q ( n ) ) ≤ 1 − α \Pr^{(\beta)}(\tau_{H(n)}>Q(n))\le1-\alpha Pr ( β ) ( τ H ( n ) > Q ( n )) ≤ 1 − α for every β ∈ B \beta\in\mathcal{B} β ∈ B simultaneously. Since A M ( x ) ≤ H ( n ) A_M(x)\le H(n) A M ( x ) ≤ H ( n ) and crossing times are monotone in the target, ( M , Q ) (M,Q) ( M , Q ) witnesses L ∈ U B , α L\in\mathcal{U}_{\mathcal{B},\alpha} L ∈ U B , α .
For the reverse inclusion, B \mathcal{B} B is nonempty; fix any β 0 ∈ B \beta_0\in\mathcal{B} β 0 ∈ B , so β 0 ≥ b ∗ > 0 \beta_0\ge b_*>0 β 0 ≥ b ∗ > 0 . A common witness for B \mathcal{B} B also witnesses membership for the single clock η β 0 \eta_{\beta_0} η β 0 , and C η β 0 , α = E X P \mathcal{C}_{\eta_{\beta_0},\alpha}=\mathbf{EXP} C η β 0 , α = EXP by Theorem 6.6, so U B , α ⊆ E X P \mathcal{U}_{\mathcal{B},\alpha}\subseteq\mathbf{EXP} U B , α ⊆ EXP . It remains to identify the leading constant. Couple the geometric waiting times of Theorem 6.11 using common iid uniforms U r , s U_{r,s} U r , s : put G r ( β ) = min { s ≥ 1 : U r , s ≤ β } G_r^{(\beta)} = \min\{s \ge1 : U_{r,s} \le\beta\} G r ( β ) = min { s ≥ 1 : U r , s ≤ β } . Then G r ( β ) G_r^{(\beta)} G r ( β ) and the cycle lengths D r ( β ) = G r ( β ) + 1 D_r^{(\beta)} = G_r^{(\beta)} + 1 D r ( β ) = G r ( β ) + 1 are nonincreasing in β > 0 \beta> 0 β > 0 . The resulting completed-heavy-sample counts N q ( β ) N_q^{(\beta)} N q ( β ) are nondecreasing in β \beta β . For each parameter this is the renewal representation in Theorem 6.11, independent of the heavy samples.
For T > 2 q T > 2q T > 2 q and q ≥ 1 q \ge1 q ≥ 1 , the lower bound in (24) and this count ordering give
L b ∗ ( T , q ) : = 1 − E ( 1 − 1 h ( T ) ) N q ( b ∗ ) ≤ inf β ∈ B Pr ( τ T ≤ q ) ≤ Pr ( τ T ≤ q ) . L_{b_*}(T,q) := 1 - \mathbb{E}\left(1 - \frac{1}{h(T)}\right)^{N_q^{(b_*)}} \le\inf_{\beta\in\mathcal{B}} \Pr(\tau_T \le q) \le\Pr(\tau_T \le q). L b ∗ ( T , q ) := 1 − E ( 1 − h ( T ) 1 ) N q ( b ∗ ) ≤ β ∈ B inf Pr ( τ T ≤ q ) ≤ Pr ( τ T ≤ q ) .
where the last inequality is Theorem F.4. For q = ⌊ c log 2 T ⌋ q = \lfloor c \log_2 T \rfloor q = ⌊ c log 2 T ⌋ with fixed c > 0 c > 0 c > 0 , the fixed-parameter renewal argument in Theorem 6.11 gives N q ( b ∗ ) / log 2 T → c / ( 1 + 1 / b ∗ ) N_q^{(b_*)}/\log_2 T \to c/(1 + 1/b_*) N q ( b ∗ ) / log 2 T → c / ( 1 + 1/ b ∗ ) in probability and h ( T ) / log 2 T → 1 h(T)/\log_2 T \to1 h ( T ) / log 2 T → 1 . For natural logarithms, the expansion N q ( b ∗ ) ln ( 1 − 1 / h ( T ) ) = − N q ( b ∗ ) / h ( T ) + o ( 1 ) N_q^{(b_*)}\ln(1 - 1/h(T)) = -N_q^{(b_*)}/h(T) + o(1) N q ( b ∗ ) ln ( 1 − 1/ h ( T )) = − N q ( b ∗ ) / h ( T ) + o ( 1 ) holds in probability, since N q ( b ∗ ) ≤ q = O ( log 2 T ) N_q^{(b_*)} \le q = O(\log_2 T) N q ( b ∗ ) ≤ q = O ( log 2 T ) . The powers are bounded by one, so the expectation argument in Theorem 6.11 applies and their expectations tend to exp ( − c / ( 1 + 1 / b ∗ ) ) \exp(-c/(1+1/b_*)) exp ( − c / ( 1 + 1/ b ∗ )) . Both bounding expressions converge to 1 − exp ( − c / ( 1 + 1 / b ∗ ) ) 1 - \exp(-c/(1 + 1/b_*)) 1 − exp ( − c / ( 1 + 1/ b ∗ )) , so the lower envelope has that limit.
Set c ∗ = ( 1 + 1 / b ∗ ) ℓ α c_* = (1 + 1/b_*)\ell_\alpha c ∗ = ( 1 + 1/ b ∗ ) ℓ α . For any 0 < ε < c ∗ 0 < \varepsilon< c_* 0 < ε < c ∗ , the limiting lower-envelope probabilities at c ∗ − ε c_* - \varepsilon c ∗ − ε and c ∗ + ε c_* + \varepsilon c ∗ + ε are respectively below and above α \alpha α . Fix error margins smaller than half these two positive gaps. For all sufficiently large T T T the lower-envelope probabilities fall below and above α \alpha α , respectively. Monotonicity in the integer budget squeezes the common quantile between the corresponding floor budgets for all sufficiently large T T T . Dividing by log 2 T \log_2 T log 2 T and letting ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 proves the limit.
References