Large-Cap Significance Paradox

• Large-Cap Significance Paradox is a phenomenon where significance assessments shift when large-scale, practically meaningful effects are isolated in high-dimensional, statistical, or financial settings.
• Analyses show that while classical spherical cap L2 discrepancy benefits from a 'blessing of dimensionality', reweighting toward large caps removes this advantage, necessitating a renormalized approach.
• Empirical studies indicate that both fixed-level hypothesis testing and market impact assessments are influenced by scale, challenging conventional interpretations of significance and risk in large-cap contexts.

Searching arXiv for the cited papers to ground the article in current records. The phrase “Large-Cap Significance Paradox” is used in the supplied literature in more than one sense. Most explicitly, it denotes the contrast on $\mathbb{S}^d$ between the classical spherical cap $L_2$ discrepancy, whose information complexity decreases with dimension, and a modified discrepancy that emphasizes caps close to hemispheres, for which the apparent high-dimensional ease disappears (Brauchart et al., 23 Apr 2026). Related papers explicitly connect the same phrase to large-sample hypothesis testing, where fixed-level significance can reject a sharp null for numerically tiny departures (Wijayatunga, 18 Mar 2025, Lovric, 28 Nov 2025), and to finance, where large-cap scale does not eliminate structured market impact or remove the large-cap leg from determining the statistical behavior of size portfolios (Vasaikar, 22 Jun 2026, Ciliberti et al., 2017). This suggests a “Large-Cap Significance Paradox” (Editor’s term) for scale-sensitive situations in which a conventional global notion of significance changes materially once large-scale, practically meaningful, or structurally central effects are isolated.

1. Spherical-cap discrepancy as the canonical formulation

On the unit sphere $$\mathbb{S}^d:=\{\boldsymbol{x}\in\mathbb{R}^{d+1}:\|\boldsymbol{x}\|_2=1\}$$, equipped with normalized surface measure $\sigma_d$, spherical caps are defined by

$$C(\boldsymbol{x};t):=\{\boldsymbol{y}\in\mathbb{S}^d:\langle \boldsymbol{x},\boldsymbol{y}\rangle\ge t\}, \qquad \boldsymbol{x}\in\mathbb{S}^d,\ t\in[-1,1].$$

Here $t=0$ gives a hemisphere, $t>0$ gives smaller caps, and $t<0$ gives larger-than-hemisphere caps. For an $N$-point set $$P_{N,d}=\{\boldsymbol{x}_1,\dots,\boldsymbol{x}_N\}\subseteq\mathbb{S}^d$$, the classical spherical cap $L_2$0 discrepancy is

$L_2$1

Its information complexity is

$L_2$2

The key theorem gives

$L_2$3

Thus, for fixed $L_2$4, the upper bound decreases as $L_2$5 increases. The paper calls this a blessing of dimensionality, in direct contrast to the usual curse-of-dimensionality paradigm. In RKHS terms, the classical discrepancy corresponds to numerical integration in $L_2$6 with kernel

$L_2$7

where

$L_2$8

Since $L_2$9, one has $$\mathbb{S}^d:=\{\boldsymbol{x}\in\mathbb{R}^{d+1}:\|\boldsymbol{x}\|_2=1\}$$0, so the associated RKHS degenerates toward the space of constants. Integration of constants is trivial; this is the analytic expression of the blessing (Brauchart et al., 23 Apr 2026).

The corresponding Stolarsky invariance principle is

$$\mathbb{S}^d:=\{\boldsymbol{x}\in\mathbb{R}^{d+1}:\|\boldsymbol{x}\|_2=1\}$$1

with

$$\mathbb{S}^d:=\{\boldsymbol{x}\in\mathbb{R}^{d+1}:\|\boldsymbol{x}\|_2=1\}$$2

and asymptotics

$$\mathbb{S}^d:=\{\boldsymbol{x}\in\mathbb{R}^{d+1}:\|\boldsymbol{x}\|_2=1\}$$3

Geometrically, the effect is driven by concentration of measure in cap volumes. For fixed $$\mathbb{S}^d:=\{\boldsymbol{x}\in\mathbb{R}^{d+1}:\|\boldsymbol{x}\|_2=1\}$$4, $$\mathbb{S}^d:=\{\boldsymbol{x}\in\mathbb{R}^{d+1}:\|\boldsymbol{x}\|_2=1\}$$5 decays exponentially in $$\mathbb{S}^d:=\{\boldsymbol{x}\in\mathbb{R}^{d+1}:\|\boldsymbol{x}\|_2=1\}$$6. Hence, in the classical average over $$\mathbb{S}^d:=\{\boldsymbol{x}\in\mathbb{R}^{d+1}:\|\boldsymbol{x}\|_2=1\}$$7, increasingly many tested caps are effectively tiny in measure, and the discrepancy increasingly concentrates around behavior near $$\mathbb{S}^d:=\{\boldsymbol{x}\in\mathbb{R}^{d+1}:\|\boldsymbol{x}\|_2=1\}$$8, i.e. near hemispheres. The paper states that “in the limit, as $$\mathbb{S}^d:=\{\boldsymbol{x}\in\mathbb{R}^{d+1}:\|\boldsymbol{x}\|_2=1\}$$9, the discrepancy measure of a point set becomes trivial (only spherical caps close to the hemisphere ($\sigma_d$0) matter in this case)” (Brauchart et al., 23 Apr 2026).

2. The balanced large-cap variant and the disappearance of the blessing

To counter the collapse of the classical model, the paper introduces a large spherical cap $\sigma_d$1 discrepancy

$\sigma_d$2

where $\sigma_d$3. When $\sigma_d$4 is small, only caps with heights in $\sigma_d$5 are tested, so one focuses on caps close to hemispheres. For $\sigma_d$6, the associated kernel satisfies

$\sigma_d$7

and the modified Stolarsky invariance principle is

$\sigma_d$8

The corresponding inverse complexity is

$\sigma_d$9

and Theorem $$C(\boldsymbol{x};t):=\{\boldsymbol{y}\in\mathbb{S}^d:\langle \boldsymbol{x},\boldsymbol{y}\rangle\ge t\}, \qquad \boldsymbol{x}\in\mathbb{S}^d,\ t\in[-1,1].$$0 yields the two-sided estimate

$$C(\boldsymbol{x};t):=\{\boldsymbol{y}\in\mathbb{S}^d:\langle \boldsymbol{x},\boldsymbol{y}\rangle\ge t\}, \qquad \boldsymbol{x}\in\mathbb{S}^d,\ t\in[-1,1].$$1

The most important choice is

$$C(\boldsymbol{x};t):=\{\boldsymbol{y}\in\mathbb{S}^d:\langle \boldsymbol{x},\boldsymbol{y}\rangle\ge t\}, \qquad \boldsymbol{x}\in\mathbb{S}^d,\ t\in[-1,1].$$2

Then

$$C(\boldsymbol{x};t):=\{\boldsymbol{y}\in\mathbb{S}^d:\langle \boldsymbol{x},\boldsymbol{y}\rangle\ge t\}, \qquad \boldsymbol{x}\in\mathbb{S}^d,\ t\in[-1,1].$$3

and therefore

$$C(\boldsymbol{x};t):=\{\boldsymbol{y}\in\mathbb{S}^d:\langle \boldsymbol{x},\boldsymbol{y}\rangle\ge t\}, \qquad \boldsymbol{x}\in\mathbb{S}^d,\ t\in[-1,1].$$4

The resulting numerical integration problem is polynomially tractable, but definitely not blessed by dimensionality. The same Sobolev smoothness class $$C(\boldsymbol{x};t):=\{\boldsymbol{y}\in\mathbb{S}^d:\langle \boldsymbol{x},\boldsymbol{y}\rangle\ge t\}, \qquad \boldsymbol{x}\in\mathbb{S}^d,\ t\in[-1,1].$$5 is retained, but the norm is renormalized so that large-cap imbalance stays significant (Brauchart et al., 23 Apr 2026).

A common misconception is that “closer to hemisphere” automatically means “harder.” The paper does not make that claim. Hemispheres are special because

$$C(\boldsymbol{x};t):=\{\boldsymbol{y}\in\mathbb{S}^d:\langle \boldsymbol{x},\boldsymbol{y}\rangle\ge t\}, \qquad \boldsymbol{x}\in\mathbb{S}^d,\ t\in[-1,1].$$6

for every $$C(\boldsymbol{x};t):=\{\boldsymbol{y}\in\mathbb{S}^d:\langle \boldsymbol{x},\boldsymbol{y}\rangle\ge t\}, \qquad \boldsymbol{x}\in\mathbb{S}^d,\ t\in[-1,1].$$7. In the limiting case $$C(\boldsymbol{x};t):=\{\boldsymbol{y}\in\mathbb{S}^d:\langle \boldsymbol{x},\boldsymbol{y}\rangle\ge t\}, \qquad \boldsymbol{x}\in\mathbb{S}^d,\ t\in[-1,1].$$8, one obtains hemisphere discrepancy with kernel

$$C(\boldsymbol{x};t):=\{\boldsymbol{y}\in\mathbb{S}^d:\langle \boldsymbol{x},\boldsymbol{y}\rangle\ge t\}, \qquad \boldsymbol{x}\in\mathbb{S}^d,\ t\in[-1,1].$$9

where $t=0$0 is normalized geodesic distance, and even $t=0$1 centrally symmetric points can make the discrepancy zero. The balanced scaling $t=0$2 is therefore not an arbitrary narrowing but a specific renormalization designed to avoid degeneration while preserving a meaningful norm (Brauchart et al., 23 Apr 2026).

A related, but distinct, cap-geometry result appears in the illumination literature. For centrally symmetric cap bodies in $t=0$3, one has

$t=0$4

and for 1-unconditionally symmetric cap bodies,

$t=0$5

This does not use the phrase “Large-Cap Significance Paradox,” but it shows that cap-focused high-dimensional geometry can remain tractable once the relevant symmetry and scale are identified (Bezdek et al., 2022).

3. Large-sample significance, Jeffreys–Lindley, and practical equivalence

In statistics, the same phrase is explicitly connected to the large-sample significance problem better known as the Jeffreys–Lindley paradox. For a sharp null hypothesis, frequentist null-hypothesis testing can reject $t=0$6 at a fixed significance level when the sample size is large, even though Bayesian posterior odds can strongly support $t=0$7, provided the null has positive prior mass and the alternative spreads its prior probability over a diffuse range of parameter values. One paper interprets this “not as an inherent incompatibility between the two inferential frameworks, but as a case where their conclusions can diverge because they are reacting to different features of the same data,” and emphasizes that it is “an instance of conflict between statistical and practical significance” arising from “using a sharp null hypothesis to approximate an acceptable small range of values for the parameter” (Wijayatunga, 18 Mar 2025).

The large-sample mechanism is explicit in Lindley’s “just significant” construction. In the normal mean problem with known variance,

$t=0$8

so that

$t=0$9

The $t>0$0-value therefore stays fixed at $t>0$1, while the raw effect size $t>0$2 shrinks to zero at rate $t>0$3. Lovric argues that this, and not prior diffuseness, is the authentic Jeffreys–Lindley paradox: it is “about more data, not vaguer priors.” In Lindley’s setup,

$t>0$4

even though the frequentist test continues to reject at fixed $t>0$5. Lovric distinguishes this from Bartlett’s Anomaly, where the posterior support for $t>0$6 goes to $t>0$7 because the variance of the prior under $t>0$8 goes to infinity with fixed data; the two regimes are mathematically different and “require different solutions” (Lovric, 28 Nov 2025).

The Bernoulli illustration in Section 2.1 of the 2025 paper makes the practical-significance issue concrete. Out of $t>0$9 Bernoulli trials, the observed success proportion is $t<0$0, extremely close to $t<0$1. But the standard error is only $t<0$2, so the deviation is many standard errors away from $t<0$3, yielding a $t<0$4-value below $t<0$5. The 99% confidence interval is $t<0$6, which excludes $t<0$7 by a tiny amount. The argument is that if values like $t<0$8, $t<0$9, and $N$0 are practically indistinguishable, then the rejection is driven by precision rather than by a meaningful substantive gap (Wijayatunga, 18 Mar 2025).

The proposed resolution is to replace point nulls by interval nulls or an “acceptable small range of values.” One formulation is

$N$1

and another is

$N$2

Under interval nulls, Bayesian and frequentist procedures can be aligned around practical equivalence rather than exact equality. This is a change of question rather than a reinterpretation of a fixed sharp-null test (Wijayatunga, 18 Mar 2025, Lovric, 28 Nov 2025).

4. Large-cap market impact and the persistence of structure in deep liquidity

In market microstructure, the paradox appears as the failure of “large-cap means too liquid for significant impact.” The paper on Apple Inc. (AAPL) tests the square-root law of market impact on a single U.S. large-capitalisation equity using the full Nasdaq TotalView-ITCH market-by-order feed over 178 trading days, from 2 December 2024 to 19 August 2025, with approximately $N$3 billion events. Metaorders are reconstructed from anonymous Level-3 order flow, and impact is calibrated as

$N$4

With the exponent fixed at $N$5, the estimated prefactor is

$N$6

and the anonymous-reconstruction correction gives

$N$7

A weighted regression with free exponent gives

$N$8

which excludes linear impact. The square-root fit has $N$9, the linear specification has $$P_{N,d}=\{\boldsymbol{x}_1,\dots,\boldsymbol{x}_N\}\subseteq\mathbb{S}^d$$0, and the Akaike difference against linear is

$$P_{N,d}=\{\boldsymbol{x}_1,\dots,\boldsymbol{x}_N\}\subseteq\mathbb{S}^d$$1

Against logarithmic impact, the square root is also preferred, with $$P_{N,d}=\{\boldsymbol{x}_1,\dots,\boldsymbol{x}_N\}\subseteq\mathbb{S}^d$$2 (Vasaikar, 22 Jun 2026).

The structural tests are equally central. In the observed data, $$P_{N,d}=\{\boldsymbol{x}_1,\dots,\boldsymbol{x}_N\}\subseteq\mathbb{S}^d$$3 of reconstructed metaorders have price moves aligned with their flow direction. Under sign shuffling, this collapses to roughly $$P_{N,d}=\{\boldsymbol{x}_1,\dots,\boldsymbol{x}_N\}\subseteq\mathbb{S}^d$$4. A permutation test centered at $$P_{N,d}=\{\boldsymbol{x}_1,\dots,\boldsymbol{x}_N\}\subseteq\mathbb{S}^d$$5 places the observed $$P_{N,d}=\{\boldsymbol{x}_1,\dots,\boldsymbol{x}_N\}\subseteq\mathbb{S}^d$$6 at $$P_{N,d}=\{\boldsymbol{x}_1,\dots,\boldsymbol{x}_N\}\subseteq\mathbb{S}^d$$7 standard deviations from the null, with

$$P_{N,d}=\{\boldsymbol{x}_1,\dots,\boldsymbol{x}_N\}\subseteq\mathbb{S}^d$$8

Under event-chronology scrambling,

$$P_{N,d}=\{\boldsymbol{x}_1,\dots,\boldsymbol{x}_N\}\subseteq\mathbb{S}^d$$9

scrambled calibrations yield a viable fit. The paper frames these as “sign and chronology fingerprints” of genuine impact rather than accidental correlation (Vasaikar, 22 Jun 2026).

The same dataset also exhibits the “two ingredients of universality theories”: long-memory order flow and diffusive prices. The aggressor-sign autocorrelation decays as

$L_2$00

while the price process remains close to diffusive, with variance ratio within $L_2$01 of $L_2$02 and Hurst exponent

$L_2$03

The prefactor is stable across 32 weekly walk-forward recalibrations, with mean $L_2$04 and range $L_2$05. The paper’s conclusion is not that large-cap liquidity removes impact, but that “even in this very liquid large-cap, market impact is not only statistically detectable, but economically structured, robust to reconstruction choices, and consistent with universality theories of impact” (Vasaikar, 22 Jun 2026).

5. Size portfolios, turnover, and the large-cap leg

A second financial use of the paradox concerns the size premium. Ciliberti, Sérié, Simon, Lempérière, and Bouchaud argue that the weakness of conventional SMB is partly an artifact of how “size” is measured and how the factor is neutralized. They define an ADV-based signal, Cold-Minus-Hot (CMH), as long low-ADV stocks and short high-ADV stocks. In their worldwide simulation, the headline long-run $L_2$06-statistics are:

• beta-neutralized SMB: $L_2$07 since 1950;
• beta-neutralized CMH: $L_2$08;
• beta- and Low-Vol-neutralized SMB: $L_2$09;
• beta- and Low-Vol-neutralized CMH: $L_2$10.

These results imply that the weak apparent size effect in standard SMB is heavily contaminated by beta and especially Low-Vol exposures. SMB is strongly anti-correlated with Low-Vol, at approximately $L_2$11. By contrast, CMH is much less anti-correlated with Low-Vol, because beta is non-monotonic in market cap but monotonic increasing in ADV, and the correlation between volatility and ADV is much weaker than the correlation between volatility and market cap (Ciliberti et al., 2017).

The paradoxical element is that the extreme portfolio risk and a substantial part of the factor’s statistical behavior are not driven by the supposedly risky small-cap leg. The paper reports that SMB portfolios are only weakly negatively skewed, CMH portfolios are virtually unskewed, and at the individual stock level small-cap and small-ADV stocks have positive skewness. More strikingly, “the extreme risk of these portfolios is dominated by the large cap leg.” On both the most positive and the most negative portfolio-return days, the biggest contributions come from the largest-ADV decile, i.e. the “hot” stocks on the short side of CMH. For market-cap SMB, the beta-neutralized large-cap leg has negative performance with a $L_2$12-stat of $L_2$13 (Ciliberti et al., 2017).

This does not mean that small stocks are riskless. The paper also shows that the probability of a $L_2$14-$L_2$15 weekly drawdown declines sharply with size, from roughly $L_2$16 for small-ADV stocks to $L_2$17 for large-ADV stocks. But the authors stress that this idiosyncratic tail risk is “clearly diversifiable,” while negative tail correlations are significantly larger among large cap stocks than among small cap stocks. The result is a mismatch between the stock-level risk story and the portfolio-level significance of the factor (Ciliberti et al., 2017).

6. Benchmark valuation, benchmark centrality, and broader interpretation

A broader benchmark-centered version appears in Knuteson’s market-microstructure hypothesis for U.S. equity valuations. The paper does not present a standard asset-pricing model with formal equilibrium derivations. Instead, it proposes a mechanical mechanism: if a sufficiently large participant repeatedly buys aggressively in the morning and sells later in the day, prices can drift upward because aggressive orders have greater price impact early in the day than later. The representative numerical example uses capital of $L_2$18B, a morning aggressive trade of $L_2$19M, with the morning full spread crossed at $L_2$20 bps and the afternoon full spread crossed at $L_2$21 bps. The expected daily trading cost is about $L_2$22B existing portfolio is about $L_2$23M (Knuteson, 2016).

The paper directly raises the question of why the S&P 500 cyclically adjusted price-to-earnings ratio has been “oddly high” for about two decades. Its back-of-the-envelope claim is that an average daily nudge of $L_2$24 basis points, sustained over seven years, can double prices: $L_2$25 The mechanism is proposed as one possible explanation for persistent overvaluation, gains to capital outpacing gains to wages, and inequality. The paper is explicit that the mechanism is testable and that, if knowingly done, the conduct would be illegal (Knuteson, 2016).

Across these domains, the shared pattern is not a single theorem but a recurring change in what counts as a meaningful test. In the spherical setting, uniform averaging over all cap heights makes high-dimensional integration look artificially easy until hemisphere-scale imbalance is reweighted (Brauchart et al., 23 Apr 2026). In hypothesis testing, fixed-$L_2$26 rejection of a sharp null can track shrinking standard errors rather than substantive effect size (Wijayatunga, 18 Mar 2025, Lovric, 28 Nov 2025). In finance, large-cap depth does not imply negligible impact, and large-cap or high-turnover legs can dominate factor-level tail behavior (Vasaikar, 22 Jun 2026, Ciliberti et al., 2017). A plausible implication is that the paradox is best understood as a warning about appropriateness of the significance notion under scale: once large-scale structure, practical equivalence, or benchmark centrality is preserved, the original appearance of ease, insignificance, or irrelevance often changes materially.