A memory is a physical system with distinguishable states. Recording, retaining, reading and resetting those states are different operations. Their thermodynamic costs depend on the complete process, including preparation, available correlations and the environment.
A blank register is a resource
A classical copying operation can be logically reversible when it keeps the input. For example, (x,y) → (x,x XOR y) copies x into y if y initially equals zero. Its inverse is the same operation. If y initially contains an unknown value, the result is not a faithful copy of x.
The blank register therefore matters. A machine that keeps writing into new blanks consumes an organized resource even if its abstract update is invertible. Returning occupied registers to a standard state lets a finite device operate repeatedly. Resetting either possible input to the same output removes the distinction from that register.
Logical irreversibility is a property of the specified input–output map. Thermodynamic entropy production is a property of its physical implementation and preparation. Confusing these questions encourages the incorrect claim that every useful calculation or every observation must release the same fixed amount of heat.
Calculate the ideal reset bound
Consider an ideal symmetric one-bit memory, initially equally likely to occupy two logical states. The memory is reset reliably to one state while in contact with a reservoir at temperature T. Assume equal initial and final mean internal energies, a control potential restored to its original form, no exploitable side information, and proper accounting for the controller.
The memory entropy decreases by kB ln 2. The reservoir must gain at least that amount of entropy on average. With Qbath positive for heat delivered to the bath, the bound is ⟨Qbath⟩ ≥ kBT ln 2. Under the stated energy conditions, the average supplied work has the same lower bound.
At 300 K, using kB = 1.380649 × 10⁻²³ J/K, this is approximately 2.871 × 10⁻²¹ J per reset. A billion independent ideal resets would have a corresponding lower bound of about 2.871 × 10⁻¹² J. These are limiting energy accounts, not estimates of the power used by an ordinary computer.
At the reversible limit, the memory's entropy decrease and the bath's entropy increase cancel. The positive heat delivered to the bath is therefore not identical to positive total entropy production. Operating irreversibly produces additional entropy. Finite-time controls and imperfect reset require their own analysis.
A stored bit need not contain one bit of uncertainty
If the initial logical probabilities are p and 1−p, the dimensionless entropy is h(p) = −p ln p − (1−p) ln(1−p). For an ideal protocol tailored to that known distribution, the corresponding reversible reset cost is kBT h(p), under the same symmetric-memory assumptions.
At p = 0.9, h(p) is about 0.325083, or 46.9% of ln 2. The physical register still has two possible logical values, but its preparation has less uncertainty. Simply counting storage cells does not determine the minimal work for a specified ensemble.
Accessible side information changes the task again. If another retained register reliably identifies the bit, a controlled reversible operation can restore the target register while keeping the distinction elsewhere. A full cycle must include what eventually happens to that other register. No cost has been eliminated merely by moving it outside the chosen boundary.
What a laboratory comparison establishes
Bérut and colleagues implemented a one-bit memory using a colloidal particle in a controlled double-well potential. Their 2012 experiment compared average heat with the Landauer bound and approached that bound for long erasure cycles. This tests a physical reset protocol under stated conditions; it does not measure a universal energy price for every act of thinking or seeing.
For a proposed device, record the initial distribution, reset success, work convention, bath temperature and all auxiliary memories. Average heat should be compared over the complete defined ensemble. Individual microscopic trials fluctuate and need not each satisfy an average inequality.
Give completed records a physical budget
The SMM/DTT proposal treats records as outcomes available after inner generation. To explain their persistence, a model must specify how recording degrees of freedom are prepared, how long they remain distinguishable and what reusing them costs. The three-arrow distinction keeps record accumulation separate from thermodynamic production.
A finite archive is a useful test. State its capacity and predict what changes when it fills: completion may stop, old distinctions may be exported, or registers may be reset. These alternatives have different observable and thermodynamic consequences. They make the physical meaning of retained history concrete.
Sources and further reading
- R. Landauer, Irreversibility and Heat Generation in the Computing Process (1961).
- A. Bérut et al., Experimental verification of Landauer’s principle linking information and thermodynamics, Nature 483, 187–189 (2012).
- Continue with computation as physical work.

