A primitive-cell NVT trajectory can have large temperature fluctuations even when its kinetic-energy distribution is correct. The same trajectory can still give a poor approximation to a bulk observable because it excludes most nuclear wavevectors. These are different questions: does the algorithm sample its finite system correctly, and does that finite system represent the material property of interest?

This note combines our existing rutile TiO₂ tests with two published studies: finite-wavevector soft modes in photoexcited silicon, and supercell restrictions on octahedral rotations in perovskites. The TiO₂ section reports calculations already completed in this research session. The literature sections use the papers' published results, with a figure for each; neither study was reproduced here.

1. What a fluctuating NVT temperature means

The temperature reported at each MD step is usually a kinetic-energy estimator. Let \(K\) be the nuclear kinetic energy, \(k_B\) the Boltzmann constant, and \(f\) the number of active momentum degrees of freedom. Define

\[ T_{\mathrm{kin}}=\frac{2K}{f k_B}. \]

The bath temperature \(T_0\) is a parameter of the canonical distribution. It is not a requirement that this instantaneous estimator remain close to \(T_0\) at every step. For classical nuclei with quadratic kinetic energy, the canonical momentum density is Gaussian. In independent mass-weighted momentum coordinates \(p_j\),

\[ K=\frac12\sum_{j=1}^{f}p_j^2, \qquad \rho(\mathbf p)\propto \exp\!\left[-\frac{\sum_j p_j^2}{2k_BT_0}\right]. \]

Consequently, the dimensionless sum of squared Gaussian variables follows a chi-square distribution:

\[ \frac{2K}{k_BT_0}=\frac{fT_{\mathrm{kin}}}{T_0}\sim\chi_f^2, \qquad \langle T_{\mathrm{kin}}\rangle=T_0, \qquad \frac{\sigma_T}{T_0}=\sqrt{\frac{2}{f}}. \]

This follows from the mean \(f\) and variance \(2f\) of the chi-square distribution. It is also the kinetic-energy distribution that a canonical thermostat must preserve; see Bussi, Donadio, and Parrinello. A correct temperature mean is necessary, but the distribution and sampling of coordinates require additional checks.

For an unconstrained periodic solid with the total center-of-mass momentum removed, \(f=3N-3\), where \(N\) is the number of atoms. Do not subtract molecular rotational degrees of freedom from a periodic solid. Other constraints change \(f\), and a program's printed convention must be checked separately.

AtomsMomentum degrees of freedomCanonical relative standard deviationStandard deviation at 300 K
61536.5%110 K
6418910.3%31 K
1624836.44%19 K

These are analytic canonical widths, not measured trajectories. A six-atom solid can therefore show large excursions without a thermostat failure. Requiring a relative standard deviation below 10% would require \(f\geq200\), or at least 68 atoms under this particular convention. That does not guarantee that every frame lies within a ±10% band, and it is not a universal minimum size for MD.

Analytic canonical kinetic-temperature probability densities: broad for six atoms and narrower for 64 and 162 atoms.
Figure 1. Analytic canonical distributions of the dimensionless ratio \(T_{\rm kin}/T_0\); the vertical axis is probability density in that ratio. Brown, blue, and green correspond to 6, 64, and 162 atoms with \(f=3N-3\). The dashed line marks the bath temperature. These are equilibrium formulas, not MD histograms.

The uncertainty of a time-averaged temperature is a different quantity. If an equilibrated trajectory supplies \(M_{\mathrm{eff}}\) effectively independent observations, its mean has an uncertainty of order \(\sigma_T/\sqrt{M_{\mathrm{eff}}}\). Correlated adjacent frames do not count as independent observations. Thermostat coupling changes the correlation time and may affect mixing; deliberately suppressing the expected kinetic-energy variance does not establish better canonical sampling. Small, nearly harmonic systems can also expose ergodicity problems in deterministic thermostats.

2. Our TiO₂ tests: which observables change?

The calculations behind this discussion used bulk rutile TiO₂: a six-atom primitive cell and its \(3\times3\times3\), 162-atom supercell. The existing VASP trajectories used a 0.5 fs timestep, ENCUT = 400 eV and EDIFF = \(10^{-6}\) eV. This section reports work already completed in the research session; no new material calculation was run for this article. The 162-atom cell is a finite reference, not an established thermodynamic limit.

The initial three-scheme comparison

The first comparison retained the inherited preparation: 0.5 ps NVT with velocity rescaling every ten steps, followed by 0.5 ps NVE. This preparation was not certified as canonical sampling. It compared primitive electronic Gamma, a primitive Gamma-centered \(3^3\) electronic mesh, and supercell electronic Gamma.

Historical NVE schemeMean T / KMean KS gap ± SD / eVMean-structure displacement RMS / Å
6 atoms, electronic Gamma204.513.36036 ± 0.121220.3671
6 atoms, electronic \(3^3\)339.141.86262 ± 0.104810.0214
162 atoms, electronic Gamma310.641.61878 ± 0.045760.0182

The last column compares each time-averaged structure with the common starting crystal after removing overall translation; it is not an instantaneous thermal displacement. The primitive Gamma trajectory moved around a substantially different mean structure. An initial-configuration check found oxygen force RMS values of about 5.18 eV/Å at primitive Gamma, 0.192 eV/Å with the primitive \(3^3\) mesh, and 0.196 eV/Å in the replicated cell. Small reconstruction and numerical-grid differences remain in this check.

The gap change also cannot be explained mainly by searching additional k points for a lower conduction edge: on the same primitive \(3^3\) trajectory, the mean Gamma direct gap exceeded the sampled-k spectral gap by only 13.99 meV. This is much smaller than the approximately 1.50 eV difference between the two primitive trajectories. Electronic integration changed the forces and the structures sampled. These short, differently heated trajectories do not establish electronic or nuclear size convergence.

Matched thermostat settings: gaps differ more than local bonds

A later test used the primitive \(3^3\) electronic mesh and supercell Gamma, each with two velocity and stochastic seeds. All four NVT inputs used the same Langevin settings: MDALGO = 3, friction 10 ps⁻¹ for Ti and O, and a 300 K target. Each seed completed 1 ps NVT followed by 1 ps NVE. The initial coordinates were shared and not already thermalized, so different seeds alone do not establish independent equilibrium samples.

The fixed NVT comparison window is steps 1001–2000, the final 0.5 ps. It was chosen by time, not by selecting frames close to 300 K, and remains a candidate sampling window rather than certified equilibrium. In this Langevin implementation the reported total kinetic temperature uses \(3N\) degrees of freedom; it is distinct from the center-of-mass-removed convention in Section 1 and the historical table above. The primitive temperature SDs are 77.12 and 88.53 K, versus 20.34 and 16.57 K for the supercell.

The gap is the sampled-k KS spectral gap, \(E_g=\min_k\epsilon_{\mathrm{CBM},k}-\max_k\epsilon_{\mathrm{VBM},k}\), at fixed electron count. Supercell folding is included. It is not an optical or quasiparticle gap. All ± values below are within-trajectory population SDs, not uncertainty bars on the means.

NVT, final 0.5 psMean T / KMean gap ± SD / eVMean Ti–O / ÅSingle-bond thermal SD / Å
6 atoms / seed 1284.251.74832 ± 0.140281.9642160.066436
6 atoms / seed 2275.441.78979 ± 0.159241.9646050.068883
162 atoms / seed 1288.661.61374 ± 0.055331.9656120.069767
162 atoms / seed 2301.151.60558 ± 0.050701.9653910.069333

Equal-weighting the two seeds, the primitive gap is higher by 0.15939 eV; the ratio of mean within-seed gap variances is 8.00. Between-seed mean differences are not included in that variance ratio. The final 0.25 ps retains a gap difference even when both primitive mean temperatures are close to 300 K, but this is still not a matched-temperature equilibrium comparison. One primitive seed's two quarter-picosecond mean gaps differ by about 0.103 eV, and the block-error diagnostic has no established plateau.

Existing TiO2 NVT gap means, variances, fixed block means and sensitivity of estimated mean errors.
Figure 2. Existing dual-seed NVT results. Red/orange are the six-atom cell; blue/teal are the 162-atom cell. Upper panels show final-0.5-ps means with SD bars and gap variances. Lower panels show fixed 125 fs block means over the full trajectory and a block-SE diagnostic. Grey shading marks the comparison window. The SE curves have no validated plateau.

Local geometry is less sensitive in this sample. The equal-seed mean Ti–O distance differs by only about 0.0011 Å, and individual-bond thermal SDs remain near 0.066–0.070 Å in both cells. This SD first removes each fixed bond's own time mean. It does not combine the static spread between long and short bonds with thermal motion.

By contrast, the temporal SD of the spatially averaged first-shell bond length is 0.00319/0.00277 Å in the primitive seeds and 0.000777/0.000449 Å in the supercell. Similar local motion and smaller fluctuations of a whole-cell average can coexist. Spatial averaging and correlations matter. The short trajectories and different measured temperatures do not certify bond-length equality to a 0.001 Å tolerance.

Similar TiO first-shell bond distributions and individual-bond fluctuations, with smaller fluctuations of the supercell mean.
Figure 3. The same NVT window and colors as Figure 2. Left: first-shell Ti–O distance probability densities in Å⁻¹, not an RDF. Right: solid bars give the RMS temporal SD of fixed individual bonds; hatched bars give the temporal SD of their instantaneous cell average. The vertical axis is logarithmic. These are different estimators of structural fluctuation.

The 153 K seed and the gap correlation

The NVE launches retained their actual NVT endpoint positions and velocities; no velocity rescaling was used to force a 300 K launch. All four 1 ps NVE segments passed their numerical energy-conservation checks. Their actual mean temperatures nevertheless differ substantially:

NVE, full 1 psMean T / KMean gap ± SD / eV\(\rho(10\,\mathrm{fs})\)First nonpositive ACF / fs
6 atoms / seed 1326.331.83932 ± 0.163680.44116.0
6 atoms / seed 2153.491.75614 ± 0.107520.43516.5
162 atoms / seed 1309.251.60237 ± 0.05164-0.01010.0
162 atoms / seed 2285.511.60042 ± 0.060410.03911.0

The primitive seed at 153.49 K is therefore not a trajectory of bulk dynamics at 300 K. Its temperature alone also does not prove that the preceding finite-cell NVT was incorrect: an NVE launch retains one sampled total energy, and kinetic and potential energies continue to exchange. Here the stronger limitation is that equilibrium sampling and the appropriate launch-energy distribution have not been established.

For each window containing \(N_s\) saved frames, the computed gap covariance subtracts that window's own mean and averages each lag over its actual available pairs:

\[ C(\ell\Delta t)=\frac{1}{N_s-\ell} \sum_{j=1}^{N_s-\ell}(E_{g,j}-\overline E_g) (E_{g,j+\ell}-\overline E_g), \qquad \rho(\tau)=\frac{C(\tau)}{C(0)}. \]

The covariance retains fluctuation amplitude; normalization compares temporal shape. At 10 fs the full and half-window primitive correlations span 0.335–0.467, versus −0.071–0.039 for the supercell. Thus the present difference is not only a difference in variance. First zero crossings describe an initial oscillation or loss of correlation; later recurrences prevent interpreting them as lifetimes.

TiO2 NVE gap covariance and normalized correlations, including sensitivity to the first and last half of a trajectory.
Figure 4. Existing NVE results, using the colors of Figures 2–3. Upper left: gap covariance in eV²; upper right: normalized ACF over 0–100 fs. Legend temperatures are actual NVE means. Lower panels compare full, first-half, and last-half windows for seed 1 of each size; each window uses its own mean. The complete 28 window curves are linked below.

The complete 0–100 fs shape is less reproducible: half-window ACF distances overlap the range of cross-cell distances. Different NVE temperatures, finite sampling and cell size remain intertwined. The defensible result is that these existing primitive trajectories do not reproduce the supercell gap statistics and short-time dynamics, while their first-shell geometry is much closer. Neither the supercell nor the primitive cell has been certified against an infinite-size reference.

3. Which missing modes matter to an observable?

Every finite supercell omits infinitely many bulk wavevectors. That fact alone does not make every finite-cell result unusable: a smooth Brillouin-zone integral can converge through increasingly fine discrete sums. The scientific question is whether the omitted spatial fluctuations carry appreciable weight for the chosen observable.

The six-atom trajectory repeats the same displacement in every translated cell. The 162-atom cell additionally permits the 27 primitive wavevectors with fractional components 0, 1/3 and 2/3. Refining electronic k integration does not introduce these nuclear degrees of freedom.

A useful local model follows when an intensive, translation-invariant observable \(A\) is smooth around a reference crystal. Translation symmetry forbids linear terms at nonzero primitive q; allowed Gamma linear terms have zero average in a centered harmonic reference. Let \(N_c\) be the number of primitive cells and \(Q_{\mathbf q\nu}\) a mass-weighted normal coordinate, with branch index \(\nu\). Use eigenvectors normalized to unit norm over the basis atoms and coordinates scaled so the harmonic potential is \(\frac12\sum_{\mathbf q\nu}\omega_{\mathbf q\nu}^2|Q_{\mathbf q\nu}|^2\). Focus on the quadratic part in the phonon eigenbasis:

\[ A^{(2)}= \frac{1}{2N_c}\sum_{\mathbf q,\nu\nu'} G_{\nu\nu'}(\mathbf q) Q_{\mathbf q\nu}^{*}Q_{\mathbf q\nu'}. \]

The matrix \(G\) is the cell-normalized curvature of the observable in these coordinates; it need not be diagonal in the phonon basis. For stable classical harmonic modes, after excluding true zero modes, the equilibrium covariance is diagonal. Define \(g_{\mathbf q\nu}=G_{\nu\nu}(\mathbf q)\). Then

\[ \langle Q_{\mathbf q\nu}^{*}Q_{\mathbf q\nu'}\rangle =\delta_{\nu\nu'}\frac{k_BT_0}{\omega_{\mathbf q\nu}^2}, \qquad \langle A^{(2)}\rangle= \frac{k_BT_0}{2N_c} \sum_{\mathbf q\nu}\frac{g_{\mathbf q\nu}}{\omega_{\mathbf q\nu}^2}. \]

This illustrative expansion explains the joint importance of mode amplitude and coupling to the observable. A soft mode may matter strongly when its coupling is appreciable. Frequency alone is insufficient, and an unstable mode cannot be assigned a harmonic thermal variance by inserting a negative \(\omega^2\) into this formula. Band crossings, large distortions, and strong anharmonicity may also invalidate the smooth expansion.

Time correlations ask for additional information. A trajectory may give an acceptable mean gap while missing the fluctuations or relaxation channels needed for the connected gap correlation

\[ C_{E_g}(t)= \left\langle [E_g(t)-\langle E_g\rangle] [E_g(0)-\langle E_g\rangle] \right\rangle. \]

The same definition, preparation, temperature or energy ensemble, and observation window must be used across cells. Missing q modes change available channels and their coupling; short trajectories add frequency-resolution and sampling errors. A velocity spectrum summed over all q is not the same observable as a common-q projected spectrum.

The appropriate reference is therefore observable-specific: the thermodynamic-limit average at fixed model, density, temperature, and electronic accuracy, or the corresponding limit of a specified dynamical correlation. For correlations, the time interval and order of size/time limits also matter. Agreement of one observable does not certify every other observable.

4. Published case: nonthermal melting of silicon

Corradini, Marini, and Calandra (2025) predict finite-q transverse-acoustic instabilities by constrained DFPT and reproduce them with a GAP potential. The authors associate the softening with a carrier-induced Kohn anomaly. At 0.2 photoexcited electrons/holes per atom and initial lattice temperature 0.1 K, 64 atoms do not melt within the reported window; 512 and 1728 do. These are ideal-lattice launches on a photoexcited surface, conserving \(K_I+F_{\rm el}\), not equilibrated NVT-to-NVE examples.

Conventional-cell repetitionAtomsReported low-temperature behavior
\(2\times2\times2\)64No melting in the observed interval.
\(4\times4\times4\)512Nonthermal melting.
\(6\times6\times6\)1728Similar nonthermal melting behavior.

In the 512-atom cell, 36 of 1536 modes are unstable. Around 700 fs their combined squared projection weight of the normalized displacement is about 97%; the 64-atom cell excludes the relevant unstable q modes. This is a group weight, not a single-q or energy fraction.

Reported silicon mode counts compared with the displacement weight in their unstable subspace: about 2.3 percent of modes carry about 97 or 87 percent of normalized squared displacement.
Figure 5. Reported scalars from Corradini et al., Fig. 12, redrawn here: blue is the unstable fraction of all modes; brown is their squared normalized displacement weight near 700 fs. The percentages have different denominators. No time traces are reconstructed.

The limit is low-temperature, classical nonthermal melting: room-temperature small cells can also melt, and later liquid motion is not confined to the initial soft subspace.

Why a small group of modes can dominate

The general mechanics behind this observation is local instability. For one mass-weighted real mode coordinate \(Q\), the harmonic equation is \(\ddot Q+\omega^2Q=0\). If \(\omega^2=-\gamma^2<0\), with growth rate \(\gamma>0\), its solution is

\[ Q(t)=Q(0)\cosh(\gamma t) +\frac{\dot Q(0)}{\gamma}\sinh(\gamma t). \]

A tiny component along an unstable direction grows exponentially while stable modes initially oscillate. Excluding that coordinate can remove the instability from a finite simulation. Negative curvature alone does not prove complete melting: nonlinear evolution determines whether the system reaches another crystal, a disordered state, or another outcome. The published trajectories and projections supply the material-specific evidence.

A quartic model \(U(x)=-ax^2+bx^4\), with \(a,b>0\), illustrates how nonlinear terms bound the motion. With effective mass \(m\), differentiating gives \(m\ddot x=2ax-4bx^3\). It explains initial growth and subsequent turnover; fitting such a model to a trajectory is a mechanism description, not an independent prediction of the whole liquid phase.

5. Published case: rotations and electronic structure

Zhao, Wang, Malyi, and Zunger (2021) explicitly discuss supercell size. Particular ordered octahedral rotations, such as \(a^0a^0b^-\), require compatible periodicity. Simple odd repetitions cannot represent those exact alternating patterns. Their Fig. 1 compares static SrTiO₃ energies across cell sizes and shows that admitting rotation patterns changes the accessible energy lowering.

Original published SrTiO3 figure: SCAN and PBEsol static energy lowering for cells from one to five repetitions, with energy-lowering ordered rotations in the tested even cells.
Figure 6. Original Fig. 1 from Zhao, Wang, Malyi, and Zunger, Materials Today 49, 107–122 (2021), reproduced unchanged from the publisher image. Horizontal axis: cell repetition; vertical axis: static energy lowering relative to the minimal-cell reference, in meV per formula unit. Red circles are SCAN; blue triangles are PBEsol. Insets show ideal and ordered tilted structures. The tested even cells admit energy-lowering rotations excluded by the tested odd cells. Lines connect discrete results; this is a published static comparison, not a new MD calculation.

The same study connects structural distortions to band gaps and band-edge structure. This establishes two linked points: a cell can exclude a relevant structural coordinate, and electronic observables depend on that coordinate. It is not a matched-temperature odd-versus-even MD benchmark. In particular, its table comparing high-symmetry, statically distorted, and thermally distorted gaps must not be read as a pure cell-size correction.

The restriction concerns the specified ordered modes. Odd cells can contain other distortions and locally tilted environments; even cells are not automatically converged. See Fig. 1 and “The choice of supercell size matters.”

A simple picture explains the commensurability constraint. Along an axis requiring an alternating rotation, the pattern is \(+,-,+,-,\ldots\). Closing an odd-length ring cannot preserve alternation across the boundary. A larger odd cell may approximate local order through defects or other patterns, but it still cannot host that exact periodic arrangement. The resulting error in an intensive observable can nevertheless shrink with size; exclusion of the exact pattern does not imply that every odd-size thermodynamic limit is wrong.

6. How to assess an MD trajectory

The TiO₂ tests and the two literature cases make the practical criterion concrete: validate the trajectory against the observable and the ensemble it is intended to represent. A temperature plot is one check within that assessment.

QuestionUseful evidenceWhat it does not establish
Are the forces numerically reliable?Electronic k-mesh, cutoff and SCF convergence; timestep sensitivity; NVE energy behavior.Nuclear size convergence or correct configurational sampling.
Does NVT sample its finite-cell equilibrium?Correct degrees of freedom; kinetic distribution and equipartition; stationary structural and energetic observables; blocked uncertainties and independent starts.Representativeness of that cell for the bulk material.
Is a thermodynamic observable converged?Cell-size and shape comparisons at matched density, temperature and electronic accuracy; changes smaller than the target tolerance with resolved uncertainty.Convergence of a different observable or a time correlation.
Is a dynamical observable converged?Matched initial ensemble; multiple launches or justified energy-shell sampling; converged timestep, duration, cell and correlation window.Exact quantum dynamics from classical nuclei, or physical lifetimes from finite-window broadening alone.

A practical sequence is to choose the target quantity and tolerance first, converge the electronic description on representative configurations, then compare several nuclear cells and adequate sampling windows. Report sampling uncertainty separately from the difference between cells. Near an apparent soft mode, inspect whether the tested cell includes its wavevector and nearby region; an atom-count sequence alone may hide a commensurability problem.

For VASP, inspect actual forces and electronic convergence rather than relying on normal termination. At the NVT-to-NVE handoff, retain the real positions and velocities and verify their units and restart semantics. Check conserved-energy behavior over a meaningful production window; a short oscillatory segment can have a misleading fitted slope. The official MD tutorial supplies the implementation context. Thermostat-generated time correlations need their own justification because a thermostat can alter dynamics even when it preserves equilibrium sampling.

Thus “primitive-cell NVT is incorrect” is usually too imprecise. A more useful conclusion names the failure: incorrect sampling of the finite Hamiltonian, insufficient statistics, numerical force/integration error, or a finite-cell bias in a specified observable relative to its bulk limit.

Sources, figures, and diagnostic code

Figure 1 is analytic. Figures 2–4 preserve results of the existing TiO₂ calculations. Figure 5 is new artwork summarizing explicitly reported silicon values; the original time traces remain in the paper. Figure 6 is the unchanged published perovskite figure. The literature studies were not reproduced, and none of these results establishes infinite-size convergence for TiO₂.