NOTE-1064: Luppi — Convergent controllers of information integration and its anaesthetic breakdown

Contribution

The paper takes the revised whole-minus-sum measure of integrated information, ΦR, and applies it to fMRI from four mammalian species under six anaesthetic regimes. In the main analysis ΦR falls under every regime but one. It rises again on recovery, and in the macaque it rises again when central thalamic stimulation restores arousal while the anaesthetic is still running. The same group’s 2024 paper (LIT-949) had shown this for human propofol and disorders of consciousness. This paper adds three more species, five more drugs, a causal reversal, and a mechanistic story from transcriptomics, connectomes and models. What it measures is the state of being responsive, at rest, with no contents and no report.

Key insight

Different anaesthetics have different molecular targets, and different species have different brains. If one change in brain dynamics appears under all of them whenever the animal stops responding, and goes away whenever it responds again, that change is a candidate for what the drugs have in common. The authors offer whole-brain pairwise ΦR as that change. They use thalamic stimulation to separate it from the mere presence of the drug: ΦR “co-varies with the consciousness-suppressing effect of anaesthetics, not with their mere presence in the system” (p. 781).

Assumptions

• Responsiveness stands in for consciousness. The target is “breakdown of interaction with the environment” (p. 781). The paper says the drugs suppress “responsiveness and presumably consciousness” (p. 781). It grants that “Behavioural unresponsiveness is an imperfect marker of unconsciousness since it can also occur as a result of sensory disconnection or motor impairment, neither of which is the same as unconsciousness” (p. 792).
• Linear-Gaussian dynamics. The decomposition uses the minimum mutual information (MMI) redundancy for Gaussian variables, on the ground that “linear-Gaussian models are sufficiently good descriptors” of fMRI (p. 795). MMI makes one source’s unique information zero by definition, and in the dynamical decomposition it “can thus lead to negative-signed atoms” (p. 795).
• Pairwise and averaged. ΦR is computed for each pair of regions and averaged over all pairs. The authors call this “perhaps best viewed as a lower bound” (p. 793). A check with 2,000 random sets of six regions, split three against three, is the only beyond-pairwise analysis (pp. 790–791).
• Intrinsic information. ΦR is the information the past of the system carries about its own future, “in the absence of any explicit tasks” (p. 778). No stimulus is presented.
• Not a test of IIT. The intention is “‘not’ to test integrated information theory (in any of its versions: 2.0, 3.0 or 4.0)” (p. 796).
• Linear control. Control energy uses the linear time-invariant model ẋ = Ax(t) + Bu(t) on each species’ structural connectome, with time horizon T = 1, between successive fMRI volumes (p. 797).

Key results

• The measure (pp. 778–779; Methods pp. 795–796). In integrated information decomposition the time-delayed mutual information of two regions splits into 16 atoms. Φ2008 = I(X_past, Y_past; X_future, Y_future) − I(X_past; X_future) − I(Y_past; Y_future). The sum of the parts counts the persistent redundancy Red→Red twice, so Φ2008 can be negative. ΦR = Φ2008 + Red→Red: the synergy-containing atoms plus the two pure-transfer atoms. In supplementary simulations, not read here, ΦR is higher for network motifs with recurrent connections, and highest for the empirical human connectome against rewired nulls (p. 796).
• The data (pp. 779–781). Humans: N = 15, awake, deep sevoflurane, and spontaneous recovery. Macaques: N = 5, scanned awake and under sevoflurane, propofol or ketamine. Marmosets: N = 4, 12 scans each, under sevoflurane, propofol or isoflurane. Mice: N = 43, awake (10), medetomidine–isoflurane (14) or halothane (19), between animals.
• ΦR falls under anaesthesia (pp. 780–781, Fig. 2). Whole-brain mean ΦR is lower under anaesthesia in every species and drug, “except for medetomidine–isoflurane in the mouse” (p. 781). Human ΦR rises again on recovery. In the six-region check both mouse regimens show a significant fall (p. 791, Extended Data Fig. 4).
• Thalamic stimulation restores it (pp. 781–782, Fig. 3). An independent macaque dataset (Tasserie et al.) had deep-brain stimulation in “N = 2 anaesthetized macaques” under propofol (p. 781). Central (centro-median) thalamic stimulation at 5 V and at 3 V raised ΦR against propofol alone. High-intensity stimulation of the ventrolateral thalamus, the control site, did not. Low-intensity ventrolateral stimulation gave “a smaller but significant increase” without restoring arousal (p. 781). Each ventrolateral condition comes from one animal.
• ΦR tracks arousal best (pp. 781–782, Fig. 3c; p. 791). A dominance analysis predicts the macaque arousal score (0 to 11) from ΦR, Φ2008, causal density and net information flow. ΦR carries 51% of the explained variance (R² = 0.57, P < 0.001 by permutation). With functional connectivity added, ΦR carries 46% and connectivity 15% (p. 791). Φ2008 is unchanged under anaesthesia in humans and macaques, rises in mice and marmosets, and goes negative in several datasets. Causal density misses recovery in humans and changes in neither primate. Net information flow behaves like ΦR and is the second-best predictor (pp. 790–791).
• Control energy rises (pp. 782–784, Fig. 4). The energy needed to move between successive brain states is higher under anaesthesia across species and drugs, “except for halothane in the mouse” (p. 782). It falls again on recovery and under central thalamic stimulation, and also partly under low-intensity ventrolateral stimulation. Across individuals it correlates negatively with ΦR in each dataset (Fig. 4g). The authors’ gloss: “the brain’s capacity to integrate information is systematically diminished when brain dynamics are less controllable” (p. 784).
• Where ΦR falls (p. 784, Fig. 5). By network-based statistic, most significant edge changes are decreases, “overwhelmingly so” in human and macaque. Most cross functional systems. “The highest prevalence of significant reductions in integrated information involves unimodal visual and somatomotor cortices” (p. 784). Default-mode and other association regions also fall, so the finding “does not contradict the importance of the default mode network (DMN) and association cortex in supporting consciousness” (p. 784). The pattern is called “widespread but unimodal dominated” (p. 792).
• PVALB (pp. 784–786, Fig. 5). Of 81 conserved receptor and cell-type genes, PVALB/Pvalb, a marker of fast-spiking inhibitory interneurons, has the most consistent spatial relation to regional ΦR loss. Against spatial-autocorrelation nulls: human rho(98) = −0.40, P = 0.01; macaque rho(80) = −0.61, P < 0.001; mouse, over regions that lose ΦR, rho(24) = −0.28, P = 0.008, 95% CI [−0.60, 0.13], and over regions that gain it, rho(44) = 0.48. Larger loss goes with more PVALB. SST is positive in humans (rho = 0.27, P = 0.045) and null in macaque and mouse. The restoration under central thalamic stimulation follows the same PVALB map (Extended Data Fig. 2). Subcortically there is no PVALB relation, so the link “may be a primarily cortical phenomenon” (p. 786).
• Thalamic core and matrix (pp. 786–787). In humans, ΦR loss is stronger in thalamic nuclei rich in matrix cells. Mouse data show no such relation, by mRNA or by protein.
• PVALB as control input (p. 787, Fig. 6). Down-weighting each region’s control input by its PVALB expression raises transition energy more than autocorrelation-matched surrogate maps do, in all four datasets tested (Hedges’ g from 0.54 to 9.82).
• Models (pp. 788–789, Figs. 7–8). Dynamic mean-field models were fitted to each species’ awake dynamics on its own connectome. Raising regional inhibition by PVALB expression lowers simulated ΦR more than null maps do, in human, macaque and mouse (g = 1.00, 1.52 and 0.70). The model is “not directly fitted to match the anaesthesia data” (p. 788). A macaque model fitted to the anaesthetized state, with thalamic nuclei that differ only in their tractography, gains more ΦR from simulated central than from ventrolateral thalamic stimulation (g = 1.07). The authors grant that ventrolateral stimulation “achieves greater success in our model than in real data” (p. 793).
• Robustness (pp. 790–791). Debiasing against surrogates, doses, another parcellation, three motion strategies and protein-based PV densities are reported to leave the results qualitatively unchanged. Some individual results lose significance with motion as a covariate.
• Discussion (pp. 791–794). Anaesthesia and brain lesions causing disorders of consciousness “may have functionally equivalent effects” (p. 791). ΦR is called “unequivocally the most consistent information-theoretic marker of anaesthesia and its reversal” (p. 792). The authors propose that anaesthesia acts on output modulation (PV) rather than input modulation (SST). They concede that “the consistency was not perfect” (p. 793). The weaker mouse effects are put down to lighter anaesthesia and different drugs.

Claims

id	claim	strength	support
C1	Whole-brain mean pairwise ΦR from resting fMRI is lower under anaesthesia than in wakefulness, across four species and five of six regimes	strong as a measured contrast	Fig. 2; pp. 780–781; six-region check, Extended Data Fig. 4
C2	ΦR returns on recovery, spontaneously in humans and under central thalamic stimulation in anaesthetized macaques	moderate: humans n = 15; stimulation in two macaques, controls from one	Figs. 2c, 3b; p. 781
C3	ΦR tracks the macaque arousal score better than Φ2008, causal density, net information flow and functional connectivity	moderate: one dataset’s behavioural scale, dominance analysis	Fig. 3c; p. 791
C4	Control energy between successive fMRI states rises under anaesthesia and correlates negatively with ΦR	moderate: linear model, uniform control, one mouse exception	Fig. 4; pp. 782–784
C5	Regional ΦR loss follows cortical PVALB expression in human, macaque and mouse	moderate: spatial correlations against autocorrelation nulls; cortical only	Fig. 5; pp. 784–786
C6	PVALB-weighted inhibition is especially suited to lowering controllability and ΦR	weak to moderate: linear-control and mean-field models, not fitted to anaesthesia data	Figs. 6–7; pp. 787–788
C7	The connectivity of the central thalamus suffices to make it a better stimulation target than the ventrolateral thalamus	weak: one model, which overstates ventrolateral success	Fig. 8; pp. 788–789, 793
C8	Anaesthesia and DOC-inducing lesions may have functionally equivalent effects	speculative	prior DOC results by the same group; p. 791

Concepts

• Integrated information decomposition (ΦID): the extension of partial information decomposition to the information a pair of variables’ past carries about their joint future, giving 4 × 4 = 16 atoms grouped as persistence, copy, transfer, erasure, upward and downward causation (p. 795).
• ΦR: whole-minus-sum integrated information that subtracts the “proper” sum of the parts, so persistent redundancy is not counted twice; equal to Φ2008 + Red→Red (pp. 779, 796).
• Φ2008: Balduzzi and Tononi’s whole-minus-sum measure, here read as the balance between ΦR and persistent redundancy.
• Transition (control) energy: the minimum summed squared input that drives the linear network model from one fMRI volume to the next, averaged over a scan (p. 797). Higher energy means less controllable dynamics.
• Responsiveness: interaction with the environment, scored behaviourally in macaques on a 0–11 arousal scale (p. 782); the paper’s operational target.

Connections

The measure is Mediano et al.’s revised integrated information. The workspace paper by the same group, LIT-949 (NOTE-799), is the paper’s ref. 45, cited for propofol and disorders of consciousness reducing ΦR between a shared set of cortical regions, “among those” that fall in the present human sevoflurane data (p. 791). The macaque stimulation data are Tasserie et al. (2022); the central-thalamus literature includes Redinbaugh et al. Integration–segregation difference (Jiang et al.) is cited as converging on the same unimodal-dominated topography. Network control follows Gu, Bassett and colleagues, and the pharmacological flattening of the energy landscape under 5HT2A agonists is the authors’ earlier work. The closure notions of LIT-1306, with Luppi and Mediano among its authors, are not used.

Bearing on the record

• THEORY-240 (Proposed; LIT-949). That THEORY’s promote_when names this paper’s design: it is strengthened “if restoring consciousness by intervention (central thalamic stimulation, for example) restores gateway ΦR”.
  • Half met. Central thalamic stimulation restores ΦR in the macaque. But the paper has no gateways or broadcasters. Its whole-brain ΦR is an average over all pairs, and its regional maps put the largest losses in unimodal visual and somatomotor cortex, not in default-mode gateways. So it does not show what THEORY-240’s claim is about, that the change is on the gateway side.
  • Its refutation clause. THEORY-240 is in trouble “if it tracks sedation or responsiveness rather than consciousness”. This paper defines its target as responsiveness. It cannot tell the two apart and says so (p. 792).
  • The sample. The human anaesthetic here is sevoflurane, not propofol. The stimulated macaques number two.
  • Proposed change (proposal only). THEORY-240’s promote_when could add that the restored ΦR must be shown regionally on the gateway side. The status stays Proposed.
• QUESTION-048 (deficit and recovery evidence). This is the kind of deficit-and-recovery evidence the question asks about, extended to four species and to a causal reversal. It confirms THEORY-091’s limit, as LIT-965 did. Its marker of consciousness is responsiveness, which is a marker of access, and the authors say that animals “cannot provide subjective reports” (p. 792). The cross-species reach is real, but it reaches responsiveness. Nothing in it separates access from experience.
• CLAIM-412 (experience constituted by interpretive organization). The exchange cited this paper for integration breaking down under anaesthesia. It does show that. But ΦR is a content-free statistic of how much information the past of a pair of regions carries about their joint future beyond what each carries alone. It is not a measure of interpretation in TERM-073’s sense: there is no distinction, context or consequence for the system in it. So the finding bears on the claim only if “interpretive organization” requires such integration, which the claim does not say. At most it shows that some integration of this kind goes with responsiveness. That is a necessary- condition reading of production views as much as of constitution.
• CLAIM-411 (most interpretation is unconscious). Integration, as A3 named it, was one of the candidate conditions. Here it falls with responsiveness, which fits a candidate. But the paper gives no case of rich integration without responsiveness, and ΦR has not been measured here between conscious and unconscious processing within a waking brain. It does not meet the claim’s defeater, which needs the unconscious cases to fall outside a stated organization.
• CLAIM-414 (four facts that come apart). The stimulation design is an intervention that separates what the ordinary trajectory cannot: low-intensity ventrolateral stimulation partly restored ΦR and controllability without restoring behaviour. That is an empirical case of a measure and its supposed behavioural effect coming apart under intervention, in the spirit of that claim. The connection is mine.
• QUESTION-254. Integration is on the question’s list of candidates. This paper gives the candidate a measured form in four species. It does not say what the measure would give for a firm or a language model, and pairwise ΦR could be computed for either. Applying it there is a test the question could ask for, and the paper does not attempt it.
• THEORY-063 (unity as one functional cluster). Its promotion test asks for “a measure of integration in the brain … computed and not asserted” that tracks unity “across waking, dreaming, anaesthesia”. This is a computed integration measure that tracks anaesthesia, but it measures the amount of pairwise integration, not whether there is one dominant cluster. It does not meet that test.
• THEORY-497 (synergy and macro features). ΦR is mostly synergy-containing atoms, so the fall under anaesthesia is a fall in the synergistic information about the system’s own future that THEORY-497 describes, relative to the same Gaussian MMI choice.
• THEORY warranted. One. THEORY-881 states C1–C3: ΦR falls with responsiveness across species and returns with it, including under thalamic stimulation. The PVALB and model results (C5–C7) bear on no record question and are left in this note.
• No instruction for machine-learning practice.

Limitations

• Responsiveness, not experience. The authors say so (p. 792). Sensory disconnection or motor impairment would give the same marker.
• Resting state, no contents. No stimulus or task was used, so nothing here concerns what is experienced or reported, only the state.
• Pairwise, linear-Gaussian, MMI. Pairwise ΦR averaged over pairs is a lower bound by the authors’ account. The six-region check uses random bipartitions, not a minimum-information partition. Results may depend on the redundancy function; other functions are said to agree, with citations.
• Small causal sample. The stimulation result rests on two macaques for central thalamus and one for each ventrolateral condition. Low-intensity ventrolateral stimulation raised ΦR without restoring arousal, which the authors explain as neural effects not yet large enough to show in behaviour (p. 793). That explanation is untested.
• One human anaesthetic. Humans had sevoflurane only. The other drugs were given to animals.
• Mouse exceptions. The main ΦR analysis misses medetomidine–isoflurane, control energy misses halothane, the PVALB correlation in the mouse runs both ways, and the thalamic result does not replicate. The authors attribute this to lighter anaesthesia; that too is untested.
• Models. The mean-field and linear-control models show that a PVALB- shaped inhibition or a central-thalamus-shaped input is sufficient in the model, not that anaesthetics act through them. The authors grant that the drugs’ molecular targets are “not all of which are directly known to influence inhibition or PV+ interneurons” (p. 792).
• Most robustness is supplementary. It was not read.
• Data access. Raw data are available from the authors “through academic collaboration” for most datasets.

Open questions

• Does ΦR fall in disconnected but conscious states, such as ketamine dreaming or covert awareness in unresponsive patients? That would separate integration from responsiveness, which this design cannot.
• Is the restored ΦR under central thalamic stimulation on the side of the default-mode gateways of LIT-949, or spread as the anaesthetic loss is?
• What does ΦR do within a waking brain between conscious and unconscious processing of the same content, the contrast CLAIM-411 turns on?
• Does a beyond-pairwise or partition-minimizing ΦR give the same topography as the pairwise average?