Feedback-Controlled Clockworks

• Feedback-controlled clockworks are systems that actively adjust timing through state-dependent feedback to concentrate irreversible events and enhance temporal precision.
• They integrate closed-loop, autonomous, and distributed architectures to manage phase synchronization, entropy reduction, and work-dissipation tradeoffs.
• Applications span thermodynamics, quantum feedback, biochemical oscillators, and digital synchronization, demonstrating practical bounds and control mechanisms.

“Feedback-controlled clockworks” is a useful umbrella for systems in which temporal regularity is not left to passive periodicity alone, but is maintained, sharpened, or redistributed by a control process that depends on state, history, or delayed signals. In thermodynamic treatments, a clockwork is the internal system whose purpose is to temporally concentrate the irreversible events that drive entropic flow; in control-theoretic treatments, feedback means that the controller’s actions depend on measurements or state estimates; in distributed and quantum settings, the same role is played by nearest-neighbor synchronization laws, buffer-mediated frequency steering, or delayed coherent feedback (Schwarzhans et al., 2020, Cao et al., 2012, Bund et al., 2023, Lall et al., 2021).

1. Conceptual scope and principal architectures

A basic distinction is between open-loop and closed-loop control. Open-loop protocols are independent of the system state, whereas closed-loop actions depend on measurements of the system. This distinction is central in the thermodynamics of feedback controlled systems, but it also reappears in distributed clocking, biochemical oscillators, and quantum delay systems (Cao et al., 2012).

The architectures collected under this heading are heterogeneous. Some are explicitly measurement-based: a controller measures a state $X_k$ or a phase and then applies a control action $C_k$. Some are internally organized feedback loops: the KaiA–KaiB–KaiC circadian oscillator, for example, relies on KaiA sequestration by phosphorylated KaiC via KaiB, but competing models assign that sequestration either a delayed negative-feedback role or a strong positive-feedback role (Golden et al., 2018). Some are autonomous rather than measurement-based: in autonomous temporal probability concentration, the clockwork is a structured system whose internal dynamics concentrate the times at which irreversible decay events occur, without external time-dependent driving (Schwarzhans et al., 2020). Others are distributed control fabrics: in PALS/GCSoC, each module has its own tunable local oscillator and neighboring modules exchange only timing information, yet neighboring modules are guaranteed to have a phase offset substantially smaller than one clock cycle (Bund et al., 2023).

A second organizing distinction is between systems that measure local phase and systems that measure global elapsed time. Minimal-clock models show that an alternator supplies local phase information, while a drop clock measures time periods of a scale global to the problem; neither architecture exhausts the possibilities, and composite clocks, cascades, and coupled clocks interpolate between them (Robu et al., 2019). This makes “feedback-controlled clockworks” a structural rather than substrate-specific category.

2. Information-theoretic and thermodynamic limits

The most general formal framework in this literature treats feedback as a resource whose effect is bounded by information. For a single actuation, if $X$ is the system state and $C$ the control action, the Shannon entropy and mutual information are

$$H(X) = - \sum_x p_X(x)\ln p_X(x), \qquad I(X;C) = H(X) - H(X|C),$$

and the information-theoretic limit of control is

$$\Delta H_{\text{closed}} \le \Delta H_{\text{open}} + I(X;C).$$

For repeated, correlated feedback operations, the entropy change at step $k$ is

$$\Delta S_k = -k_B I(X_k;C_k \mid C_{k-1},\ldots,C_1),$$

so the total average entropy reduction due to information used in $M$ control steps is

$$\Delta S_{\text{info}} = -k_B \sum_{k=1}^{M} I(C_k;X_k \mid C_{k-1},\ldots,C_1).$$

Only non-redundant information reduces entropy; correlated measurements do not contribute their raw information content step by step (Cao et al., 2012).

This has a direct clock interpretation. If the clock macrostate is phase, or phase plus amplitude, repeated phase corrections cannot reduce uncertainty by more than the conditional mutual information in the control record. Frequent measurements can therefore be strongly redundant. This suggests that for clock stabilization the relevant quantity is not a naive “bits per measurement” count, but a conditional information rate.

The same framework links information to work and dissipation. In the state-of-the-art reviewed alongside these entropy formulas, generalized second-law-like inequalities take the form

$C_k$0

A feedback-controlled clock can reduce its internal uncertainty, but that reduction must be paid for by work and/or entropy production elsewhere, including memory erasure in the controller (Cao et al., 2012).

Concrete performance bounds are known for paradigmatic cyclic devices such as feedback flashing ratchets. If $C_k$1 is the information per control step, then

$C_k$2

These are flux and power bounds rather than clock-precision bounds, but they show that performance improvements can scale sublinearly or linearly with information, not arbitrarily. A plausible implication is that explicit timing-accuracy bounds, when derived, should also depend on non-redundant information and dissipation rather than on raw sampling frequency alone (Cao et al., 2012).

3. Minimal, biological, and biochemical clock models

Minimal-clock models make the informational structure of clockworks explicit. Time is treated as a random variable $C_k$3 with uniform prior on $C_k$4, clock states are Markovian variables $C_k$5, and clock quality is measured by the mutual information $C_k$6. The two primitive one-bit clocks are the alternator and the drop clock. The alternator is a two-state Markov chain with switch probability $C_k$7; at $C_k$8 it is a perfect alternator and gives one bit of local phase information. The drop clock is a one-way decay process with

$C_k$9

and gives coarse information about elapsed time on the scale of the whole run. Bags of clocks, cascades, and composite clocks with controlled dependency enlarge this design space. In the composite clock, the lower clock is driven by the upper one under a transfer-entropy constraint $X$0; at $X$1 bit/tick the optimal solution is a perfect 2-bit counter (Robu et al., 2019).

Biological oscillators add environmental noise and evolved feedback motifs. A limit-cycle oscillator is not essential for knowing time: bacteria can possess an hourglass, a system that when forced by an oscillatory light input exhibits robust oscillations but in the absence of driving relaxes to a stable fixed point. In models of the Kai system, a limit-cycle oscillator, an exponentially relaxing hourglass, and an oscillatory-relaxation hourglass are equally informative on time in the limit of low input-noise, yet in the regime of high input-noise the limit-cycle oscillator is far superior; the same behavior is found in the Stuart–Landau model (Monti et al., 2017). This directly separates phase-robust clockworks from purely driven ones.

The KaiA–KaiB–KaiC clock sharpens the issue by contrasting feedback architectures. In the allosteric and two-site allosteric models, KaiA sequestration acts as delayed negative feedback: highly phosphorylated inactive KaiC binds KaiB, KaiBC complexes sequester KaiA, and reduced free KaiA delays the next phosphorylation wave. In the monomer model, sequestration acts effectively as positive feedback because S-state KaiC sequesters KaiA and thereby shifts the D$X$2S balance further toward S, producing a relaxation-oscillator mechanism. The two mechanisms make opposite predictions for how period changes with sequestration strength, and they can be distinguished experimentally by introducing a protein that binds competitively with KaiA (Golden et al., 2018).

4. Autonomous and quantum clockworks

An autonomous thermodynamic clockwork is built around irreversible events rather than explicit measurements. In autonomous temporal probability concentration, the clockwork is designed to concentrate the probability that a ladder system occupies its top level into narrow time windows, so that an irreversible decay from that level produces regular ticks. If the top-level population is $X$3 and the decay-channel coupling is $X$4, the tick-time density is

$X$5

Clock quality is measured by the resolution

$X$6

and the accuracy

$X$7

The paper shows that a perfect clockwork can be approximated arbitrarily well by increasing its complexity, but the irreversible decay mechanism still imposes the ultimate thermodynamic limit to the measurement of time (Schwarzhans et al., 2020).

Quantum feedback-controlled clockworks introduce delayed coherent feedback directly into the oscillator dynamics. In a linear delayed quantum self-sustained oscillator, the expectation value obeys a delay differential equation of the form

$X$8

For suitable gain and delay, this system exhibits perfect oscillation without phase diffusion. The delayed feedback plays the role of a phase-selective memory, and the mean quadratures trace closed cycles in phase space. The price is unbounded energy growth. When nonlinear two-photon absorption is added to clamp the amplitude, the mean field dephases and the oscillator behaves like previously studied non-delayed quantum limit-cycle systems: bounded energy is recovered, but phase diffusion returns (Liu et al., 2023).

A common misconception is that autonomous and feedback-controlled clockworks are mutually exclusive. The literature does not support that dichotomy. Autonomous temporal concentration uses internal dynamics to shape tick statistics, while delayed quantum feedback uses coherent self-coupling to stabilize phase. Both are clockworks in the precise sense that they structure the timing of irreversible or observable events, even when no external measurement record is used (Schwarzhans et al., 2020, Liu et al., 2023).

5. Distributed synchronization and communication fabrics

Distributed electronic clockworks replace a single global metronome by a network of local oscillators under feedback. In PALS/GCSoC, each node $X$9 has a hardware clock $C$0 with drift $C$1, and a logical clock

$C$2

where $C$3 selects slow or fast mode. Neighbor-offset estimates drive a gradient clock synchronization rule. Under $C$4 and $C$5, the theorem quoted in the paper gives

$C$6

For a 15 nm ASIC simulation running at 2 GHz, the reported parameters yield worst-case bounds of 20 ps on the phase offset for a $C$7 node grid network (Bund et al., 2023).

Bittide pursues a different target: not wall-clock agreement, but a perfectly synchronized logical clock. In the abstract frame model, node phases satisfy

$C$8

while buffer occupancy on link $C$9 is

$$H(X) = - \sum_x p_X(x)\ln p_X(x), \qquad I(X;C) = H(X) - H(X|C),$$0

A decentralized controller such as

$$H(X) = - \sum_x p_X(x)\ln p_X(x), \qquad I(X;C) = H(X) - H(X|C),$$1

steers frequencies so that buffers neither overflow nor underflow. The paper proves existence and uniqueness of solutions under a minimum-frequency condition and emphasizes that the goal is a perfectly shared logical time rather than close alignment to UTC (Lall et al., 2021).

At the link level, a synchronizer-free digital controller implements an even tighter feedback clockwork. A producer and a consumer run on independent clocks, a ring buffer mediates communication, and a controller slightly adjusts the two oscillator frequencies according to buffer fill. Metastability is not eliminated; it is contained. The controller output may become metastable, but communication is guaranteed to remain metastability-free, and with the 65 nm implementation running at roughly 2 GHz the formal analysis permits a buffer of size $$H(X) = - \sum_x p_X(x)\ln p_X(x), \qquad I(X;C) = H(X) - H(X|C),$$2 while deterministically avoiding underrun or overflow (Bund et al., 2020).

These distributed results show that feedback-controlled clockworks need not be centralized and need not track physical wall time. They can instead enforce local skew bounds, logical rounds, or metastability-free communication by regulating phase differences and buffer states.

6. Event-time shaping, computation, and temporal resource allocation

Some clockworks act directly on event timing rather than on oscillator phase. In feedback control of waiting times, the controlled object is the waiting time distribution

$$H(X) = - \sum_x p_X(x)\ln p_X(x), \qquad I(X;C) = H(X) - H(X|C),$$3

Choosing $$H(X) = - \sum_x p_X(x)\ln p_X(x), \qquad I(X;C) = H(X) - H(X|C),$$4 after each jump is equivalent to shaping the inter-event statistics of the next tick. A Gamma target yields

$$H(X) = - \sum_x p_X(x)\ln p_X(x), \qquad I(X;C) = H(X) - H(X|C),$$5

with

$$H(X) = - \sum_x p_X(x)\ln p_X(x), \qquad I(X;C) = H(X) - H(X|C),$$6

As $$H(X) = - \sum_x p_X(x)\ln p_X(x), \qquad I(X;C) = H(X) - H(X|C),$$7, the clock approaches deterministic waiting times. The same paper contrasts active feedback with passive, hardwired clockworks and uses optimal control functionals to trade WTD matching against control effort (Brandes et al., 2016).

A broader computational reinterpretation treats clocks and timing structures as the computational fabric itself. In computing with clocks, a value is represented by a time interval

$$H(X) = - \sum_x p_X(x)\ln p_X(x), \qquad I(X;C) = H(X) - H(X|C),$$8

and unary temporal computation follows directly: addition is implemented by concatenation,

$$H(X) = - \sum_x p_X(x)\ln p_X(x), \qquad I(X;C) = H(X) - H(X|C),$$9

while multiplication is implemented by clock scaling. Digital, analog, and photonic accumulators all convert time intervals into counts or stored charge. The model is explicitly asynchronous in control, and correctness depends on stability of the delay within the data interval and on agreement about the time reference used to measure it (Edwards et al., 2024).

A recent quantum-classical extension treats runtime calibration itself as a state-trajectory feedback-control problem. The state includes an equivalent calibration age $$\Delta H_{\text{closed}} \le \Delta H_{\text{open}} + I(X;C).$$0, an optimization gap $$\Delta H_{\text{closed}} \le \Delta H_{\text{open}} + I(X;C).$$1, and the remaining wall-clock budget. Calibration quality is

$$\Delta H_{\text{closed}} \le \Delta H_{\text{open}} + I(X;C).$$2

recovery actions are costly partial resets, and policies are evaluated by the time-integrated optimization gap over the full execution window. Using a finite-horizon rollout controller, the reported ordering is clear: cloud-like feedback is generally uncompetitive, while local-ms and tight-loop regimes open a positive-gain region that grows with workload quality-sensitivity and initial calibration age; the advantage of tight-loop integration becomes pronounced under capacity pressure when many calibration targets must be processed within the same control window (Deng, 12 May 2026).

Across these domains, the common principle is that a clockwork is not merely a source of periodic motion. It is a controlled temporal structure that allocates, sharpens, or synchronizes time itself—whether by reducing entropy with measurements, concentrating irreversible events, stabilizing phase through delayed feedback, coordinating distributed oscillators, or directly shaping event intervals and wall-clock expenditure.