# Adversarial Review: Observable-Cost Result

Date: 2026-09-10

Role: adversarial reviewer. I tried to break C1-C5, with emphasis on code defects, finite-size artifacts, statistical overclaiming, and prior art.

## Executive Verdict

I did **not** find a fatal implementation bug in the core TEBD canonical-center logic, the palindromic Trotter step, the column-stacked Liouvillian, the Lindblad dissipator, or `expect_rho`. Those attacks mostly fail.

I did find several serious scope failures. The strongest are:

- **SERIOUS:** Tests A/B do not prove "observables are cheaper only by a factor of about 2." They prove that one particular compression procedure, sequential Schmidt/MPS truncation of the exact pure state, gives only a small saving for one small observable set on `n=16`, `t<=4`. A simulator optimized directly for local reduced density matrices, Heisenberg-picture operators, hydrodynamic variables, DMT/DAOE/TDVP-style effective dynamics, or sampling histories is not tested.
- **SERIOUS:** The finite-size/light-cone claim is wrong for the actual observable set. `CENTRE = 4..12` in `n=16`, not the central half; site 12 is only 3 bonds from the right edge, so with `v=2` boundary influence reaches it at `t ~= 1.5`, not `t=4`. The fitted window `[0.25,4]` is therefore not protected from boundaries.
- **SERIOUS:** Test C's late-time decline is physically real for this Lindblad model but mostly demonstrates relaxation toward a high-entropy, near-featureless mixed state. At `gamma=0.5`, purity is `0.0011` at `t=6`, almost the maximally mixed value `1/2^10 = 0.0009766`; the cheap state is cheap partly because almost all local signals have been erased.
- **SERIOUS:** C3 is largely a reproduction/rediscovery of known results, not a new finding. The prior-art thread already records Noh, Jiang & Fefferman, arXiv:2003.13163, as showing bounded maximum MPO entanglement under noise, with cost growing as error rate falls; it also records DAOE as explicitly producing operator-entanglement peak-and-decay under dissipation.
- **FATAL to the numerical power-law claim:** `chi_peak = 5.0 * gamma^(-1.25)` is not evidence. It is a two-parameter fit to three points over one decade of `gamma`, with small-`n` finite-size ceilings nearby. Treat it as a plotting annotation only.

## Code Attacks

### TEBD truncation and orthogonality center: attack failed

Severity: **MINOR / survives**.

In `mps.py`, `right_canonicalize` sweeps right-to-left and leaves the orthogonality center at site 0 (`mps.py:68-79`). In `tebd_step`, the right sweep starts at bond 0 and after `apply_gate(..., centre_moves_right=True)` stores `U` on site `i` and `s*Vh` on site `i+1`, so the center moves to `i+1` (`mps.py:133-135`, `mps.py:152-155`). The subsequent left sweep starts at `n-2` with the center on `i+1`, then stores `U*s` on site `i` and `Vh` on `i+1`, moving the center left (`mps.py:136-138`, `mps.py:155-156`). That is internally consistent.

Caveat: this is optimal only for the two-site SVD split in the current canonical gauge. It is not a global variational optimum over all MPS of bond dimension `chi`, and it is definitely not an optimum for local-observable error. Therefore it validates TEBD as implemented, but not the broad simulator-cost conclusion.

### Trotter decomposition: attack failed

Severity: **MINOR / survives**.

The step is palindromic:

```text
prod_{i=0..n-2} exp(-i h_i dt/2) prod_{i=n-2..0} exp(-i h_i dt/2)
```

That is a time-symmetric product formula and is genuinely second order (`mps.py:142-157`). It is not the usual odd/even two-layer TEBD presentation, but it is a valid second-order sweep formula. The validation claim that state infidelity falls by about 16x per halving of `dt` is also compatible with second-order state-vector error, because infidelity is quadratic in state error.

### Exact-state truncation in Test A is not the optimum observable representation

Severity: **SERIOUS**.

Test A does this:

```python
trunc = M.mps_to_vec(M.vec_to_mps(v, N, chi_max=chi))
oerr.append(max_abs_observable_error(trunc))
```

See `run_ab.py:135-140`. That asks: if I take the exact wavefunction and apply left-to-right SVD compression, what local observable error results? It does **not** ask: what is the minimum `chi` representation sufficient for the requested observables?

For observable prediction, the natural object is the local reduced density matrices or Heisenberg-evolved local operators, not necessarily a globally compressed pure state. The prior-art thread itself lists DMT, DAOE, TDVP-after-barrier, MPDO, OSEE, thermal-state, and hydrodynamic methods that intentionally give up global fidelity while preserving targeted local data. So C1's phrase "not orders of magnitude" is too broad. The result only supports "not orders of magnitude within sequential pure-state MPS truncation on this observable set and time window."

### Liouvillian, vectorization, dissipator, and expectation values: attacks failed

Severity: **MINOR / survives**.

Given column stacking, `vec(A X B) = (B^T kron A) vec(X)`. `run_c.py` uses

```python
L = -1j * (kron(Id, H) - kron(H.T, Id))
```

for `-i(H rho - rho H)` (`run_c.py:67-70`). That is correct. The dissipator uses `gamma * (kron(Zi.T, Zi) - kron(Id, Id))` for `Z rho Z - rho` (`run_c.py:71-74`). Since `Tr(Z rho Z) = Tr(rho)`, the dissipator is trace-preserving. `expect_rho` computes `sum_ij O_ij rho_ji = Tr(O rho)` using sparse COO indices (`run_c.py:127-130`), which is correct.

### Operator-space grouping in Test C: likely correct, but narrower than claimed

Severity: **MINOR for code, SERIOUS for interpretation**.

`operator_svd` reshapes `rho` into row/ket bits followed by column/bra bits, then permutes to `(ket0,bra0,ket1,bra1,...)` before the middle cut (`run_c.py:94-103`). That is the right site grouping for operator-space entanglement if the dense matrix ordering matches the Kronecker construction, which it appears to.

But Test C computes only the **middle-cut** operator SVD, then calls the resulting cutoff `chi_MPO` (`run_c.py:24-29`, `run_c.py:160-170`). A real MPO bond dimension is the maximum over all cuts, or at least a full-chain truncation scheme. Middle cut is a useful diagnostic, not a full MPO-cost measurement. For symmetric quenches the middle cut is probably the worst cut at intermediate times, but with open boundaries, dissipation, and a restricted observable set, that assumption needs to be checked rather than named as `chi_MPO`.

### Test C observable set mismatch

Severity: **SERIOUS**.

Tests A/B use single-site `Z_i`, `X_i` and two-site `Z_iZ_j`, `X_iX_j` at separations 1-4 (`run_ab.py:74-85`). Test C uses single-site `Z_i`, `X_i` and only `Z_iZ_j` at separations 1-2; it omits `X_iX_j` entirely and omits separations 3-4 (`run_c.py:113-122`).

This matters because C3 compares closed and open cases as if the observable metric is the same. It is not. Test C's peak `chi` values may be underestimates relative to the A/B observable set.

## Finite Size

Severity: **SERIOUS**.

The finite-size protection statement is wrong in the code comments. `run_ab.py` says the observable set is "central half" and that the boundary reaches the centre at `t = 4` (`run_ab.py:67-71`). But for `N=16`, `CENTRE = range(4, 13)`, i.e. sites `4..12`, which is 9 sites, not half the chain. The rightmost included site, 12, is only 3 bonds from the right boundary at site 15; with the stated `v = 2 min(1,g) = 2`, the boundary light cone reaches it at `t ~= 3/2 = 1.5`. The leftmost included site 4 is reached at `t ~= 2.0`.

The two-site observables do not fix this. They include pairs ending at site 12 and starting at site 4. Since the error metric is a **maximum** over the observable set (`run_ab.py:140`, `run_ab.py:167`), one boundary-contaminated observable can determine `chi_obs`.

For the exact middle cut, the quasiparticle saturation time is closer to the usual order `L/(2v) ~= 16/4 = 4`, so the entanglement entropy can still look plausibly pre-saturation through much of the window. But local observable boundary contamination begins much earlier than `t=4`. The central restriction is therefore partly cosmetic.

Bias direction: finite size usually caps entanglement growth and should bias late-time `alpha` **downward** once saturation begins. Boundary reflections can also create nonmonotone local observable errors, making `chi_obs` artificially easier or harder at individual times. In these logs the middle-cut entanglement keeps rising through `t=4` in the chaotic case (`fine.log:51-59`), so a pure saturation cap is not obvious by `t=4`; the stronger objection is that the observable set is already boundary-aware after `t~1.5-2`.

## Chaotic Ratio Trend

Severity: **SERIOUS, but the trend itself mostly survives**.

The reported late-window chaotic ratio slope, `+0.214 +- 0.034` for `t>=2`, is reproducible from the printed table. Using the nine ratios from `fine.log:51-59`,

```text
t:        2.00  2.25  2.50  2.75  3.00  3.25  3.50  3.75  4.00
ratio:    1.31  1.57  1.62  1.67  1.73  2.05  2.08  2.06  2.00
log-slope 0.2143 per time unit, OLS SE 0.0340
```

I tested the obvious quantization objection by treating each reported crossing as an interval between the previous chi-grid point and the reported point. A Monte Carlo over those intervals gave a positive slope in all sampled cases, with approximate 2.5/50/97.5 percentiles `0.131 / 0.179 / 0.227`. Leave-one-out slopes ranged from `0.177` to `0.253`. A simple HAC/Newey-West lag-1 standard error was about `0.0345`, essentially unchanged.

So: the claim "the ratio rises over this table" survives. What does **not** survive is using that as asymptotic evidence. The ratio only moves from about 1.3 to about 2.0 over a boundary-contaminated, small-`n`, 9-point late window. The finite chi ceiling at 256 is not the issue here; all chaotic fine-grid values are far below 256 (`fine.log:51-65`). The issue is external validity, not arithmetic.

## Test B

Severity: **SERIOUS**.

The quoted Test B sequence for chaotic `eps=1e-3`,

```text
chi=4 -> 0.25, chi=8 -> 1.25, chi=16 -> 2.25, chi=32 -> 3.25, chi=64 -> >4.0
```

is consistent with `t_max` increasing by about 1 per doubling over the measured range. But there are only five useful `chi` values, the last is right-censored at `>4.0`, and the `t` resolution is 0.25. This is evidence for logarithmic behavior in this small window, not a decisive scaling law. Also, because the observable set includes sites hit by boundary light cones after `t~1.5-2`, most of the sequence used to infer the slope is not clean thermodynamic-limit data.

The code's `t_max` definition also requires the error to stay below threshold at **every previous measurement** and stops at the first failure (`run_ab.py:172-184`). That is a reasonable operational definition, but in a small finite chain with oscillatory errors it can turn a transient spike into a permanent failure. It is stricter than "accurate at time `t`" and should be described that way.

## Test C: Physical Meaning of the Decline

Severity: **SERIOUS**.

The late-time decline is physically meaningful for the model, but it is not meaningful in the way C3 wants.

The model has dephasing in the `Z` basis plus a Hamiltonian with `X` terms. This drives the finite system toward a very mixed state. The purity values show it directly:

- `gamma=0.5`: purity falls from `1.0` to `0.0011` by `t=6` (`testc.log:54-68`). The maximally mixed value for `n=10` is `1/1024 = 0.0009766`.
- `gamma=0.2`: purity is `0.0037` by `t=6` (`testc.log:38-52`), only about 3.8 times maximally mixed.
- `gamma=0.05`: purity is still `0.0622` at `t=6` (`testc.log:22-36`), so this case is not yet fully trivial, but it is strongly mixed.

So I commit to this answer: the cost decline is a real consequence of the open-system channel, but the late-time cheapness mostly reflects erasure/thermalization toward a feature-poor density matrix. That does **not** model a universe whose observers retain structured memories, records, long-range correlations, low-entropy subsystems, and nontrivial local expectation values. It models a subsystem after the environment has destroyed much of the information in that subsystem.

This does not kill the statement "dephasing can bound operator-space cost." It kills the stronger inference "therefore inhabited open systems are cheaply simulable after the hump." The latter requires showing that observer-relevant records remain recoverable and mutually consistent while discarded coherences stay irrelevant.

## Power Law

Severity: **FATAL to the fitted law**.

The power law uses exactly three data points for two fit parameters:

```text
gamma=0.05 -> 215
gamma=0.2  -> 38
gamma=0.5  -> 12
```

That leaves one degree of freedom. The exponent `-1.25` is therefore not a measured scaling law. It is compatible with many curved crossovers, finite-size effects, threshold artifacts, and observable-set artifacts. It spans only one decade in `gamma`, on `n=10`, with a finite middle-cut ceiling of `4^5 = 1024`. The `gamma=0` case already reaches 609 by `t=6` and is near that ceiling for tighter tolerance (`testc.log:6-20`).

At most, Test C supports the qualitative statement "larger dephasing lowered the observed peak cost in this small model." It does not support `chi_peak ~ gamma^-1.25`, and not even "roughly `1/gamma`" as more than a hunch.

## C4: Discarding the Environment / Garbage Collection

Severity: **SERIOUS**.

C4 is conceptually plausible but overstates what the experiment proves.

In ordinary open-system simulation, tracing out the environment is allowed only when no future observer can access records that purify the reduced state. But a universe-scale simulator has to keep observers consistent with each other. If an environmental record can later be brought into causal contact with an observer, interfered, decoded, or used in a quantum eraser protocol, discarding it is not free. The simulator either has to retain enough purification data to answer those future correlations, or impose real physical irreversibility that prevents those questions from being asked.

Quantum erasure is the sharp counterexample. Dephasing looks irreversible in the reduced density matrix, but if the environment is coherently controlled, apparent which-path information can be erased and interference restored. In that case "discard what the environment learned" gives wrong predictions. C4 only works under an extra condition: environmental degrees of freedom must become permanently inaccessible, thermodynamically scrambled, or objectively collapsed such that no future experiment can recombine the branches.

The garbage-collection analogy is therefore conditional, not established by Test C. Test C uses a Lindblad semigroup where irreversibility is inserted by hand. It does not show that the full closed system containing environment plus observers is cheap.

## C5: Objective-Collapse Connection

Severity: **SERIOUS**.

C5 is an interesting bridge but the analogy is too loose as stated.

The experiment uses site-local dephasing in the lattice `Z` basis:

```text
gamma sum_i (Z_i rho Z_i - rho)
```

See `run_c.py:19-20` and `run_c.py:71-74`. GRW/CSL-style objective collapse is normally formulated as stochastic localization in position/mass-density space, not arbitrary dephasing in the computational `Z` basis of a spin-chain simulator. Diosi-Penrose is tied to gravity/mass distribution. A simulator-imposed decoherence floor would need a preferred basis, a rate law, a scaling with mass/particle number/energy, and noise correlations. Test C supplies none of those.

The right restrained statement is: if simulation-cost pressure produces real, irreducible decoherence in observable degrees of freedom, then collapse/decoherence experiments are the natural place to constrain it. The wrong statement is: Test C meaningfully points to GRW/CSL/DP as the same signature. Basis and rate structure matter.

## Prior Art / Novelty

Severity: **SERIOUS**.

Argus is rediscovering known results for C3, not opening a new domain.

The prior-art thread already says Noh, Jiang & Fefferman, arXiv:2003.13163, use MPO entanglement entropy as a cost proxy and find that above a characteristic size, "the maximum achievable MPO entanglement entropy is bounded by a constant that depends only on the gate error rate, not on the system size," with cost increasing as error rate decreases (`observable-cost-prior-art.md:209-211`). That is the core of C3.

The same prior-art thread also says DAOE applies artificial dissipation to Heisenberg-picture operators and that "the dissipation leads to a decay of operator entanglement," with nonzero dissipation causing operator entanglement to peak and decrease while permitting finite-`chi` long-time simulations (`observable-cost-prior-art.md:141-147`). That is even closer to Argus's "finite hump" story.

Prosen & Znidaric are also directly relevant. The prior-art thread records their 2007 result that non-integrable tilted-field Ising dynamics has `D_epsilon(t) ~ exp(h_q t)` with `h_q = 1.10`, independent of error, initial observable, and `n` once `n` is large enough, while integrable cases grow polynomially or saturate depending on operator class (`observable-cost-prior-art.md:97-99`). This already contains the closed-system operator-space growth contrast and makes Argus's integrable/non-integrable small-chain comparison less novel.

So the novelty of Argus's Test C is only a local reproduction on his exact toy model and code. That is useful for calibration, but C3 should be framed as "I reproduced the known noisy/open-system OSEE hump in this model," not as an independent discovery.

## Attacks by Conclusion

### C1. Local observables are cheaper only by a factor of about 2; exponential character unchanged.

Severity: **SERIOUS**.

The measured factor-of-2 statement is too broad. It is true for the logged `eps=1e-3` fine-grid pure-state truncations by `t=4`: integrable ratio peaks around 1.8 and ends at 1.42 (`fine.log:16-24`), chaotic ratio reaches about 2.0 (`fine.log:51-59`). But that is one compression family and one finite observable set. Prior-art methods exist precisely because local observables can be captured after global fidelity fails. C1 must be narrowed to: "within sequential pure-state MPS truncation, for this small closed chain and observable set, the saving is modest through `t<=4`."

### C2. Closed-system accessible time grows only logarithmically in resources; H6b should go up.

Severity: **SERIOUS**.

The log-time relation is supported by standard theory for global MPS fidelity and by prior operator-space results in non-integrable systems, but this experiment alone is too small to move H6b upward. Test B has five `chi` values, one censored endpoint, `t` resolution 0.25, and boundary contamination for the observables after `t~1.5-2`. Also, worst-case local observable prediction remains BQP-hard in general while special physical regimes can be easier; neither maps cleanly to "physics permits no cheaper simulation" without defining the simulator's allowed approximation target.

### C3. Open finite dephasing gives a finite hump; H6b true for closed, false for open.

Severity: **SERIOUS**.

The finite hump survives as a small-model observation. The general conclusion does not. It is known prior art, uses a narrower Test C observable set, measures a middle-cut SVD rather than full MPO bond dimension, and late-time cheapness coincides with strong mixing toward near-maximally-mixed states. The correct conclusion is conditional: Markovian dephasing in this small chain suppresses middle-cut OSEE and the measured local-observable truncation cost.

### C4. Decoherence is garbage collection that demonstrably saves exponentially.

Severity: **SERIOUS**.

The experiment demonstrates savings only after a Lindblad trace-out has been assumed. It does not show that a simulator of a closed universe may discard environmental records while preserving all future observer correlations. Quantum erasure and recoherence are direct threats. C4 needs the condition "records discarded are permanently inaccessible in principle, or objective collapse is real."

### C5. Simulator decoherence floor connects to objective-collapse experiments.

Severity: **SERIOUS**.

The experimental-program connection is plausible but underspecified. Lindblad `Z`-dephasing on a spin chain is not GRW/CSL/DP. To make C5 more than analogy, Argus needs a preferred-basis story and a rate/noise law that can be compared with collapse bounds.

## WHAT SURVIVES

- **C1 survives only narrowly:** for sequential pure-state MPS truncation on `n=16`, `t<=4`, and the logged central-observable set, local observables were only modestly cheaper than global fidelity. Caveat: not an optimized observable-only simulator and not thermodynamic-limit evidence.
- **C2 survives as prior-supported background, not as a new strong update:** closed non-integrable 1D dynamics is expected to have exponential operator/MPS cost and `t_max ~ log chi` in many settings. Caveat: this experiment's finite-size and small-sample limits are too strong for a large credence move.
- **C3 survives qualitatively:** finite dephasing can create an OSEE/cost hump and late decline. Caveat: known prior art, small `n`, middle-cut only, narrower observables, and late-time mixed-state trivialization.
- **C4 does not survive without a new condition:** garbage collection saves only if environmental records are permanently inaccessible or physically collapsed. Otherwise the purification may need to be kept.
- **C5 survives only as a research lead:** collapse/decoherence experiments are relevant if simulator-cost decoherence is real, but this Lindblad toy model is not yet a GRW/CSL/DP signature.