Supplemantary for
Fair Streaming Principal Component Analysis: Statistical and Algorithmic Viewpoint

Junghyun Lee^∗ Hanseul Cho^∗ Se-Young Yun Chulhee Yun
Kim Jaechul Graduate School of AI, KAIST

Dataset = COMPAS [feature dim $=11$ , #(train data)=4,316]
$k$	Method	%Var( $\uparrow$ )	MMD²( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )
$k$	Method	%Var( $\uparrow$ )	MMD²( $\downarrow$ )	kernel SVM		linear SVM		MLP
2	PCA	2.09_(0.11)	0.057_(0.006)	64.84_(0.78)	0.28_(0.03)	58.91_(0.52)	0.2_(0.03)	64.16_(1.01)	0.3_(0.03)
	Olfat & Aswani, 2019 (0.1)	Memory Out
	Olfat & Aswani, 2019 (0.0)	Memory Out
	Lee et al., 2022 (1e-3)	1.78_(0.11)	0.003_(0.002)	62.65_(1.54)	0.03_(0.03)	57.56_(2.06)	0.04_(0.03)	60.79_(1.59)	0.03_(0.01)
	Lee et al., 2022 (1e-6)	1.78_(0.11)	0.003_(0.002)	62.65_(1.54)	0.03_(0.03)	57.56_(2.06)	0.04_(0.03)	61.09_(1.53)	0.03_(0.01)
	Kleindessner et al., 2023 (mean)	1.97_(0.1)	0.008_(0.003)	61.94_(0.59)	0.07_(0.02)	56.03_(1.09)	0.05_(0.04)	60.79_(0.8)	0.07_(0.04)
	Kleindessner et al., 2023 (0.85)	1.78_(0.13)	0.005_(0.002)	60.64_(1.08)	0.08_(0.05)	56.16_(1.21)	0.01_(0.02)	58.8_(1.74)	0.07_(0.04)
	Kleindessner et al., 2023 (0.5)	1.62_(0.09)	0.004_(0.001)	59.67_(0.89)	0.1_(0.03)	54.81_(1.04)	0.0_(0.0)	57.11_(1.35)	0.08_(0.05)
	Kleindessner et al., 2023 (kernel)	N/A	N/A	57.8_(1.82)	0.08_(0.06)	54.74_(1.21)	0.02_(0.04)	55.91_(1.27)	0.02_(0.03)
	Ravfogel et al., 2020	0.62_(0.18)	0.0_(0.0)	56.35_(0.71)	0.01_(0.01)	54.81_(1.04)	0.0_(0.0)	55.38_(0.66)	0.01_(0.01)
	Ravfogel et al., 2022	0.49_(0.03)	0.002_(0.002)	57.54_(0.74)	0.03_(0.03)	54.81_(1.04)	0.0_(0.0)	56.42_(1.29)	0.03_(0.03)
	Samadi et al., 2018	N/A	N/A	64.19_(1.07)	0.15_(0.03)	59.15_(0.59)	0.13_(0.03)	64.4_(1.32)	0.15_(0.03)
	Ours (offline, mean)	1.97_(0.1)	0.008_(0.003)	61.94_(0.59)	0.07_(0.02)	56.03_(1.09)	0.05_(0.04)	60.61_(1.5)	0.06_(0.05)
	Ours (FNPM, mean)	1.97_(0.1)	0.008_(0.003)	61.94_(0.59)	0.07_(0.02)	56.03_(1.09)	0.05_(0.04)	60.5_(1.48)	0.07_(0.05)
	Ours (offline, $m$ =2)	1.92_(0.11)	0.006_(0.002)	61.22_(0.85)	0.11_(0.04)	55.43_(0.9)	0.04_(0.04)	59.53_(1.35)	0.12_(0.06)
	Ours (FNPM, $m$ =2)	1.92_(0.11)	0.006_(0.002)	61.08_(0.9)	0.09_(0.04)	55.51_(1.16)	0.03_(0.04)	59.56_(1.48)	0.12_(0.07)
	Ours (offline, $m$ =5)	1.89_(0.1)	0.006_(0.002)	61.71_(0.78)	0.1_(0.04)	55.49_(0.79)	0.03_(0.04)	59.74_(2.0)	0.14_(0.06)
	Ours (FNPM, $m$ =5)	1.89_(0.1)	0.006_(0.002)	61.67_(0.76)	0.11_(0.05)	55.55_(0.9)	0.03_(0.04)	59.63_(1.84)	0.13_(0.07)
Dataset = German Credit [feature dim $=59$ , #(train data)=700]
$k$	Method	%Var( $\uparrow$ )	MMD²( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )
$k$	Method	%Var( $\uparrow$ )	MMD²( $\downarrow$ )	kernel SVM		linear SVM		MLP
2	PCA	11.13_(0.32)	0.293_(0.054)	75.47_(0.67)	0.15_(0.09)	70.3_(1.59)	0.03_(0.08)	71.43_(0.98)	0.15_(0.09)
	Olfat & Aswani, 2019 (0.1)	7.43_(0.56)	0.017_(0.009)	72.17_(1.04)	0.03_(0.02)	69.8_(1.21)	0.0_(0.0)	0.0_(0.0)	0.0_(0.0)
	Olfat & Aswani, 2019 (0.0)	7.33_(0.54)	0.015_(0.01)	71.77_(1.52)	0.03_(0.02)	69.8_(1.21)	0.0_(0.0)	0.0_(0.0)	0.0_(0.0)
	Lee et al., 2022 (1e-3)	8.89_(0.5)	0.027_(0.015)	73.23_(2.27)	0.03_(0.03)	69.97_(1.93)	0.0_(0.0)	71.6_(1.58)	0.05_(0.04)
	Lee et al., 2022 (1e-6)	8.89_(0.5)	0.027_(0.015)	73.23_(2.27)	0.03_(0.03)	69.97_(1.93)	0.0_(0.0)	70.73_(1.98)	0.03_(0.04)
	Kleindessner et al., 2023 (mean)	10.39_(0.62)	0.031_(0.018)	76.77_(1.47)	0.04_(0.03)	69.97_(1.93)	0.0_(0.0)	72.13_(1.69)	0.05_(0.03)
	Kleindessner et al., 2023 (0.85)	6.97_(0.41)	0.021_(0.008)	73.17_(2.03)	0.03_(0.03)	69.97_(1.93)	0.0_(0.0)	70.7_(2.65)	0.02_(0.02)
	Kleindessner et al., 2023 (0.5)	4.66_(0.22)	0.017_(0.013)	71.6_(2.84)	0.02_(0.02)	69.97_(1.93)	0.0_(0.0)	71.17_(2.95)	0.01_(0.01)
	Kleindessner et al., 2023 (kernel)	N/A	N/A	69.8_(1.21)	0.0_(0.0)	69.8_(1.21)	0.0_(0.0)	69.97_(1.93)	0.0_(0.0)
	Ravfogel et al., 2020	3.25_(0.38)	0.007_(0.004)	71.33_(2.32)	0.02_(0.02)	69.97_(1.93)	0.0_(0.0)	70.13_(2.03)	0.01_(0.01)
	Ravfogel et al., 2022	3.19_(0.36)	0.04_(0.024)	71.37_(2.5)	0.03_(0.04)	69.97_(1.93)	0.0_(0.0)	70.47_(2.06)	0.02_(0.03)
	Samadi et al., 2018	N/A	N/A	74.2_(2.15)	0.06_(0.04)	69.93_(1.99)	0.0_(0.01)	76.57_(2.6)	0.08_(0.06)
	Ours (offline, mean)	10.39_(0.62)	0.031_(0.018)	76.77_(1.47)	0.04_(0.03)	69.97_(1.93)	0.0_(0.0)	71.8_(1.59)	0.06_(0.05)
	Ours (FNPM, mean)	10.39_(0.62)	0.031_(0.018)	76.7_(1.43)	0.04_(0.03)	69.97_(1.93)	0.0_(0.0)	72.63_(1.78)	0.08_(0.05)
	Ours (offline, $m$ =10)	6.36_(0.51)	0.017_(0.008)	72.6_(2.37)	0.03_(0.02)	69.97_(1.93)	0.0_(0.0)	71.2_(2.14)	0.02_(0.03)
	Ours (FNPM, $m$ =10)	6.55_(0.44)	0.019_(0.01)	73.0_(2.36)	0.02_(0.02)	69.97_(1.93)	0.0_(0.0)	70.83_(2.25)	0.02_(0.02)
	Ours (offline, $m$ =25)	4.89_(0.24)	0.032_(0.018)	72.77_(2.57)	0.04_(0.04)	69.97_(1.93)	0.0_(0.0)	71.27_(2.72)	0.03_(0.04)
	Ours (FNPM, $m$ =25)	4.91_(0.29)	0.028_(0.018)	72.53_(2.37)	0.03_(0.02)	69.97_(1.93)	0.0_(0.0)	70.9_(2.33)	0.02_(0.02)
Dataset = Adult Income [feature dim $=102$ , #(train data)=31,655]
$k$	Method	%Var( $\uparrow$ )	MMD²( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )
$k$	Method	%Var( $\uparrow$ )	MMD²( $\downarrow$ )	kernel SVM		linear SVM		MLP
10	PCA	20.82_(0.38)	0.195_(0.002)	88.58_(0.21)	0.18_(0.01)	83.14_(0.21)	0.19_(0.0)	83.88_(0.23)	0.2_(0.01)
	Olfat & Aswani, 2019 (0.1)	Memory Out
	Olfat & Aswani, 2019 (0.0)	Memory Out
	Lee et al., 2022 (1e-3)	Takes too long time
	Lee et al., 2022 (1e-6)	Takes too long time
	Kleindessner et al., 2023 (mean)	18.84_(0.35)	0.01_(0.0)	88.43_(0.18)	0.18_(0.01)	81.71_(0.29)	0.03_(0.01)	83.83_(0.32)	0.17_(0.01)
	Kleindessner et al., 2023 (0.85)	14.54_(0.23)	0.002_(0.0)	86.99_(0.32)	0.16_(0.0)	75.33_(0.19)	0.0_(0.0)	81.92_(0.56)	0.14_(0.01)
	Kleindessner et al., 2023 (0.5)	11.34_(0.2)	0.0_(0.0)	83.01_(0.28)	0.12_(0.01)	75.33_(0.19)	0.0_(0.0)	79.13_(0.77)	0.08_(0.02)
	Kleindessner et al., 2023 (kernel)	Takes too long time
	Ravfogel et al., 2020	9.36_(0.26)	0.001_(0.0)	85.71_(0.78)	0.14_(0.02)	75.33_(0.19)	0.0_(0.0)	80.18_(0.68)	0.1_(0.02)
	Ravfogel et al., 2022	9.6_(0.25)	0.001_(0.0)	87.61_(0.57)	0.16_(0.01)	75.37_(0.22)	0.0_(0.0)	80.72_(0.9)	0.11_(0.02)
	Samadi et al., 2018	N/A	N/A	85.69_(0.24)	0.18_(0.01)	83.25_(0.18)	0.17_(0.01)	84.52_(0.29)	0.19_(0.01)
	Ours (offline, mean)	18.84_(0.35)	0.01_(0.0)	88.43_(0.18)	0.18_(0.01)	81.71_(0.29)	0.03_(0.01)	83.7_(0.24)	0.17_(0.01)
	Ours (FNPM, mean)	18.83_(0.35)	0.009_(0.0)	88.53_(0.22)	0.18_(0.01)	81.7_(0.29)	0.03_(0.01)	83.85_(0.28)	0.18_(0.01)
	Ours (offline, $m$ =15)	14.48_(0.22)	0.002_(0.0)	86.86_(0.3)	0.16_(0.0)	75.33_(0.19)	0.0_(0.0)	81.84_(0.54)	0.13_(0.01)
	Ours (FNPM, $m$ =15)	14.38_(0.2)	0.001_(0.0)	86.64_(0.29)	0.16_(0.0)	75.33_(0.19)	0.0_(0.0)	80.96_(0.56)	0.12_(0.02)
	Ours (offline, $m$ =50)	11.34_(0.2)	0.0_(0.0)	83.0_(0.25)	0.12_(0.01)	75.33_(0.19)	0.0_(0.0)	78.94_(0.7)	0.07_(0.02)
	Ours (FNPM, $m$ =50)	11.34_(0.2)	0.0_(0.0)	82.85_(0.22)	0.11_(0.01)	75.33_(0.19)	0.0_(0.0)	78.59_(0.63)	0.06_(0.02)

1 Synthetic Experiments: Verifying FNPM is PAFO-Learnble

Refer to caption — Figure 1: ( $\epsilon_{1}$ , $\epsilon_{2}$ , $\delta$ )-PAFO learnability. The slope $\approx-0.5$ of the log-log plot indicates the sample complexity (measured by block size) is about ${\mathcal{O}}({\epsilon}_{1}^{-2}+{\epsilon}_{2}^{-2})$ (with probability $1-\delta=0.9$ ).

2 Numerical Experiments: Fair PCA on UCI datasets

Table 1: Dataset = Adult Income [feature dim=102, #(train data)=31,655]

$k$	Method	%Var( $\uparrow$ )	MMD²( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )
$k$	Method	%Var( $\uparrow$ )	MMD²( $\downarrow$ )	kernel SVM		linear SVM		MLP
2	PCA	6.88_(0.14)	0.374_(0.006)	82.4_(0.2)	0.19_(0.01)	81.23_(0.16)	0.2_(0.0)	82.31_(0.18)	0.19_(0.01)
	Olfat & Aswani, 2019 (0.1)	Memory Out
	Olfat & Aswani, 2019 (0.0)	Memory Out
	Lee et al., 2022 (1e-3)	5.68_(0.11)	0.0_(0.0)	80.34_(0.24)	0.05_(0.01)	78.09_(0.36)	0.01_(0.0)	80.31_(0.26)	0.06_(0.01)
	Lee et al., 2022 (1e-6)	5.42_(0.11)	0.0_(0.0)	79.41_(0.23)	0.02_(0.01)	77.52_(0.69)	0.01_(0.0)	79.48_(0.29)	0.02_(0.01)
	Kleindessner et al., 2023 (mean)	5.74_(0.11)	0.002_(0.0)	80.6_(0.2)	0.07_(0.01)	78.57_(0.26)	0.01_(0.0)	80.42_(0.17)	0.07_(0.01)
	Kleindessner et al., 2023 (0.85)	4.09_(0.17)	0.001_(0.001)	75.52_(0.21)	0.0_(0.0)	75.33_(0.19)	0.0_(0.0)	75.37_(0.23)	0.0_(0.0)
	Kleindessner et al., 2023 (0.5)	2.63_(0.07)	0.0_(0.0)	75.38_(0.18)	0.0_(0.0)	75.33_(0.19)	0.0_(0.0)	75.33_(0.19)	0.0_(0.0)
	Kleindessner et al., 2023 (kernel)	Takes too long time
	Ravfogel et al., 2020	1.91_(0.08)	0.001_(0.001)	75.67_(0.31)	0.0_(0.0)	75.33_(0.19)	0.0_(0.0)	75.54_(0.39)	0.0_(0.01)
	Ravfogel et al., 2022	1.91_(0.09)	0.006_(0.011)	75.59_(0.34)	0.0_(0.0)	75.33_(0.19)	0.0_(0.0)	75.48_(0.35)	0.0_(0.0)
	Samadi et al., 2018	N/A	N/A	82.63_(0.18)	0.15_(0.01)	82.21_(0.2)	0.17_(0.0)	77.62_(3.46)	0.06_(0.09)
	Ours (offline, mean)	5.74_(0.11)	0.002_(0.0)	80.6_(0.2)	0.07_(0.01)	78.57_(0.26)	0.01_(0.0)	80.31_(0.22)	0.07_(0.01)
	Ours (FNPM, mean)	5.74_(0.11)	0.002_(0.0)	80.6_(0.2)	0.07_(0.01)	78.57_(0.26)	0.01_(0.0)	80.33_(0.25)	0.07_(0.01)
	Ours (offline, $m$ =15)	4.04_(0.14)	0.001_(0.001)	75.51_(0.23)	0.0_(0.0)	75.33_(0.19)	0.0_(0.0)	75.39_(0.24)	0.0_(0.0)
	Ours (FNPM, $m$ =15)	4.07_(0.13)	0.001_(0.0)	75.54_(0.25)	0.0_(0.0)	75.33_(0.19)	0.0_(0.0)	75.58_(0.26)	0.01_(0.01)
	Ours (offline, $m$ =50)	2.63_(0.07)	0.0_(0.0)	75.38_(0.18)	0.0_(0.0)	75.33_(0.19)	0.0_(0.0)	75.39_(0.25)	0.0_(0.01)
	Ours (FNPM, $m$ =50)	2.64_(0.06)	0.0_(0.0)	75.44_(0.21)	0.0_(0.0)	75.33_(0.19)	0.0_(0.0)	75.37_(0.2)	0.0_(0.0)
10	PCA	20.82_(0.38)	0.195_(0.002)	88.58_(0.21)	0.18_(0.01)	83.14_(0.21)	0.19_(0.0)	83.88_(0.23)	0.2_(0.01)
	Olfat & Aswani, 2019 (0.1)	Memory Out
	Olfat & Aswani, 2019 (0.0)	Memory Out
	Lee et al., 2022 (1e-3)	Takes too long time
	Lee et al., 2022 (1e-6)	Takes too long time
	Kleindessner et al., 2023 (mean)	18.84_(0.35)	0.01_(0.0)	88.43_(0.18)	0.18_(0.01)	81.71_(0.29)	0.03_(0.01)	83.83_(0.32)	0.17_(0.01)
	Kleindessner et al., 2023 (0.85)	14.54_(0.23)	0.002_(0.0)	86.99_(0.32)	0.16_(0.0)	75.33_(0.19)	0.0_(0.0)	81.92_(0.56)	0.14_(0.01)
	Kleindessner et al., 2023 (0.5)	11.34_(0.2)	0.0_(0.0)	83.01_(0.28)	0.12_(0.01)	75.33_(0.19)	0.0_(0.0)	79.13_(0.77)	0.08_(0.02)
	Kleindessner et al., 2023 (kernel)	Takes too long time
	Ravfogel et al., 2020	9.36_(0.26)	0.001_(0.0)	85.71_(0.78)	0.14_(0.02)	75.33_(0.19)	0.0_(0.0)	80.18_(0.68)	0.1_(0.02)
	Ravfogel et al., 2022	9.6_(0.25)	0.001_(0.0)	87.61_(0.57)	0.16_(0.01)	75.37_(0.22)	0.0_(0.0)	80.72_(0.9)	0.11_(0.02)
	Samadi et al., 2018	N/A	N/A	85.69_(0.24)	0.18_(0.01)	83.25_(0.18)	0.17_(0.01)	84.52_(0.29)	0.19_(0.01)
	Ours (offline, mean)	18.84_(0.35)	0.01_(0.0)	88.43_(0.18)	0.18_(0.01)	81.71_(0.29)	0.03_(0.01)	83.7_(0.24)	0.17_(0.01)
	Ours (FNPM, mean)	18.83_(0.35)	0.009_(0.0)	88.53_(0.22)	0.18_(0.01)	81.7_(0.29)	0.03_(0.01)	83.85_(0.28)	0.18_(0.01)
	Ours (offline, $m$ =15)	14.48_(0.22)	0.002_(0.0)	86.86_(0.3)	0.16_(0.0)	75.33_(0.19)	0.0_(0.0)	81.84_(0.54)	0.13_(0.01)
	Ours (FNPM, $m$ =15)	14.38_(0.2)	0.001_(0.0)	86.64_(0.29)	0.16_(0.0)	75.33_(0.19)	0.0_(0.0)	80.96_(0.56)	0.12_(0.02)
	Ours (offline, $m$ =50)	11.34_(0.2)	0.0_(0.0)	83.0_(0.25)	0.12_(0.01)	75.33_(0.19)	0.0_(0.0)	78.94_(0.7)	0.07_(0.02)
	Ours (FNPM, $m$ =50)	11.34_(0.2)	0.0_(0.0)	82.85_(0.22)	0.11_(0.01)	75.33_(0.19)	0.0_(0.0)	78.59_(0.63)	0.06_(0.02)

Table 2: Dataset = COMPAS [feature dim

=11

, #(train data)=4,316]

$k$	Method	%Var( $\uparrow$ )	MMD²( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )
$k$	Method	%Var( $\uparrow$ )	MMD²( $\downarrow$ )	kernel SVM		linear SVM		MLP
2	PCA	2.09_(0.11)	0.057_(0.006)	64.84_(0.78)	0.28_(0.03)	58.91_(0.52)	0.2_(0.03)	64.16_(1.01)	0.3_(0.03)
	Olfat & Aswani, 2019 (0.1)	Memory Out
	Olfat & Aswani, 2019 (0.0)	Memory Out
	Lee et al., 2022 (1e-3)	1.8_(0.1)	0.003_(0.002)	62.41_(1.09)	0.03_(0.02)	57.57_(1.18)	0.04_(0.03)	60.79_(1.59)	0.03_(0.01)
	Lee et al., 2022 (1e-6)	1.78_(0.11)	0.003_(0.002)	62.65_(1.54)	0.03_(0.03)	57.56_(2.06)	0.04_(0.03)	61.09_(1.53)	0.03_(0.01)
	Kleindessner et al., 2023 (mean)	1.97_(0.1)	0.008_(0.003)	61.94_(0.59)	0.07_(0.02)	56.03_(1.09)	0.05_(0.04)	60.79_(0.8)	0.07_(0.04)
	Kleindessner et al., 2023 (0.85)	1.78_(0.13)	0.005_(0.002)	60.64_(1.08)	0.08_(0.05)	56.16_(1.21)	0.01_(0.02)	58.8_(1.74)	0.07_(0.04)
	Kleindessner et al., 2023 (0.5)	1.62_(0.09)	0.004_(0.001)	59.67_(0.89)	0.1_(0.03)	54.81_(1.04)	0.0_(0.0)	57.11_(1.35)	0.08_(0.05)
	Kleindessner et al., 2023 (kernel)	N/A	N/A	57.8_(1.82)	0.08_(0.06)	54.74_(1.21)	0.02_(0.04)	55.91_(1.27)	0.02_(0.03)
	Ravfogel et al., 2020	0.62_(0.18)	0.0_(0.0)	56.35_(0.71)	0.01_(0.01)	54.81_(1.04)	0.0_(0.0)	55.38_(0.66)	0.01_(0.01)
	Ravfogel et al., 2022	0.49_(0.03)	0.002_(0.002)	57.54_(0.74)	0.03_(0.03)	54.81_(1.04)	0.0_(0.0)	56.42_(1.29)	0.03_(0.03)
	Samadi et al., 2018	N/A	N/A	64.19_(1.07)	0.15_(0.03)	59.15_(0.59)	0.13_(0.03)	64.4_(1.32)	0.15_(0.03)
	Ours (offline, mean)	1.97_(0.1)	0.008_(0.003)	61.94_(0.59)	0.07_(0.02)	56.03_(1.09)	0.05_(0.04)	60.61_(1.5)	0.06_(0.05)
	Ours (FNPM, mean)	1.97_(0.1)	0.008_(0.003)	61.94_(0.59)	0.07_(0.02)	56.03_(1.09)	0.05_(0.04)	60.5_(1.48)	0.07_(0.05)
	Ours (offline, $m$ =2)	1.92_(0.11)	0.006_(0.002)	61.22_(0.85)	0.11_(0.04)	55.43_(0.9)	0.04_(0.04)	59.53_(1.35)	0.12_(0.06)
	Ours (FNPM, $m$ =2)	1.92_(0.11)	0.006_(0.002)	61.08_(0.9)	0.09_(0.04)	55.51_(1.16)	0.03_(0.04)	59.56_(1.48)	0.12_(0.07)
	Ours (offline, $m$ =5)	1.89_(0.1)	0.006_(0.002)	61.71_(0.78)	0.1_(0.04)	55.49_(0.79)	0.03_(0.04)	59.74_(2.0)	0.14_(0.06)
	Ours (FNPM, $m$ =5)	1.89_(0.1)	0.006_(0.002)	61.67_(0.76)	0.11_(0.05)	55.55_(0.9)	0.03_(0.04)	59.63_(1.84)	0.13_(0.07)
10	PCA	5.85_(0.36)	0.089_(0.002)	82.78_(0.43)	0.15_(0.04)	65.39_(1.15)	0.16_(0.05)	70.01_(0.93)	0.2_(0.04)
	Olfat & Aswani, 2019 (0.1)	Memory Out
	Olfat & Aswani, 2019 (0.0)	Memory Out
	Lee et al., 2022 (1e-3)	5.36_(0.31)	0.002_(0.001)	80.76_(0.46)	0.1_(0.03)	63.89_(1.15)	0.04_(0.03)	68.97_(1.46)	0.03_(0.02)
	Lee et al., 2022 (1e-6)	4.83_(0.45)	0.002_(0.001)	78.93_(0.84)	0.09_(0.03)	61.68_(1.54)	0.04_(0.03)	67.23_(1.66)	0.04_(0.02)
	Kleindessner et al., 2023 (mean)	5.67_(0.37)	0.004_(0.001)	80.87_(0.54)	0.09_(0.03)	64.76_(1.04)	0.02_(0.02)	69.83_(0.83)	0.04_(0.03)
	Kleindessner et al., 2023 (0.85)	5.41_(0.33)	0.004_(0.001)	80.64_(0.82)	0.09_(0.04)	63.63_(1.19)	0.02_(0.01)	68.69_(1.28)	0.05_(0.04)
	Kleindessner et al., 2023 (0.5)	5.29_(0.31)	0.003_(0.001)	79.22_(0.44)	0.1_(0.03)	61.62_(0.63)	0.02_(0.02)	68.63_(0.92)	0.07_(0.04)
	Kleindessner et al., 2023 (kernel)	N/A	N/A	65.96_(1.12)	0.26_(0.07)	64.35_(0.8)	0.05_(0.04)	64.93_(1.49)	0.04_(0.03)
	Ravfogel et al., 2020	2.64_(0.47)	0.001_(0.0)	64.17_(1.46)	0.04_(0.02)	54.85_(1.03)	0.0_(0.0)	62.56_(1.79)	0.04_(0.03)
	Ravfogel et al., 2022	2.45_(0.06)	0.001_(0.0)	71.75_(0.69)	0.07_(0.04)	55.26_(1.25)	0.0_(0.0)	66.66_(1.08)	0.06_(0.03)
	Samadi et al., 2018	N/A	N/A	69.75_(1.0)	0.17_(0.02)	65.54_(0.72)	0.14_(0.03)	69.97_(0.81)	0.16_(0.03)
	Ours (offline, mean)	5.67_(0.37)	0.004_(0.001)	80.87_(0.54)	0.09_(0.03)	64.78_(1.04)	0.02_(0.02)	69.16_(0.5)	0.04_(0.01)
	Ours (FNPM, mean)	5.68_(0.37)	0.004_(0.001)	80.87_(0.49)	0.09_(0.03)	64.84_(1.01)	0.02_(0.02)	68.86_(0.82)	0.04_(0.03)
	Ours (offline, $m$ =2)	5.57_(0.35)	0.004_(0.001)	80.93_(0.63)	0.09_(0.03)	64.72_(0.93)	0.02_(0.02)	69.25_(1.27)	0.04_(0.04)
	Ours (FNPM, $m$ =2)	5.58_(0.35)	0.004_(0.001)	80.82_(0.48)	0.09_(0.03)	64.65_(1.1)	0.02_(0.02)	68.97_(1.31)	0.04_(0.02)
	Ours (offline, $m$ =5)	5.55_(0.35)	0.004_(0.001)	81.03_(0.61)	0.09_(0.03)	64.47_(1.01)	0.02_(0.02)	68.92_(1.05)	0.03_(0.02)
	Ours (FNPM, $m$ =5)	5.55_(0.34)	0.004_(0.001)	80.9_(0.49)	0.09_(0.03)	64.54_(1.23)	0.02_(0.02)	69.05_(1.09)	0.04_(0.03)

Table 3: Dataset = German Credit [feature dim=59, #(train data)=700]

$k$	Method	%Var( $\uparrow$ )	MMD²( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )	%Acc( $\uparrow$ )	$\Delta_{\rm DP}$ ( $\downarrow$ )
$k$	Method	%Var( $\uparrow$ )	MMD²( $\downarrow$ )	kernel SVM		linear SVM		MLP
2	PCA	11.13_(0.32)	0.293_(0.054)	75.47_(0.67)	0.15_(0.09)	70.3_(1.59)	0.03_(0.08)	71.43_(0.98)	0.15_(0.09)
	Olfat & Aswani, 2019 (0.1)	7.25_(0.65)	0.024_(0.01)	71.53_(1.85)	0.02_(0.02)	69.97_(1.93)	0.0_(0.0)	0.0_(0.0)	0.0_(0.0)
	Olfat & Aswani, 2019 (0.0)	7.28_(0.52)	0.022_(0.009)	72.03_(1.99)	0.02_(0.02)	69.97_(1.93)	0.0_(0.0)	0.0_(0.0)	0.0_(0.0)
	Lee et al., 2022 (1e-3)	9.82_(0.45)	0.027_(0.018)	74.87_(1.76)	0.05_(0.04)	69.97_(1.93)	0.0_(0.0)	71.6_(1.58)	0.05_(0.04)
	Lee et al., 2022 (1e-6)	8.89_(0.5)	0.027_(0.015)	73.23_(2.27)	0.03_(0.03)	69.97_(1.93)	0.0_(0.0)	70.73_(1.98)	0.03_(0.04)
	Kleindessner et al., 2023 (mean)	10.39_(0.62)	0.031_(0.018)	76.77_(1.47)	0.04_(0.03)	69.97_(1.93)	0.0_(0.0)	72.13_(1.69)	0.05_(0.03)
	Kleindessner et al., 2023 (0.85)	6.97_(0.41)	0.021_(0.008)	73.17_(2.03)	0.03_(0.03)	69.97_(1.93)	0.0_(0.0)	70.7_(2.65)	0.02_(0.02)
	Kleindessner et al., 2023 (0.5)	4.66_(0.22)	0.017_(0.013)	71.6_(2.84)	0.02_(0.02)	69.97_(1.93)	0.0_(0.0)	71.17_(2.95)	0.01_(0.01)
	Kleindessner et al., 2023 (kernel)	N/A	N/A	69.8_(1.21)	0.0_(0.0)	69.8_(1.21)	0.0_(0.0)	69.97_(1.93)	0.0_(0.0)
	Ravfogel et al., 2020	3.25_(0.38)	0.007_(0.004)	71.33_(2.32)	0.02_(0.02)	69.97_(1.93)	0.0_(0.0)	70.13_(2.03)	0.01_(0.01)
	Ravfogel et al., 2022	3.19_(0.36)	0.04_(0.024)	71.37_(2.5)	0.03_(0.04)	69.97_(1.93)	0.0_(0.0)	70.47_(2.06)	0.02_(0.03)
	Samadi et al., 2018	N/A	N/A	74.2_(2.15)	0.06_(0.04)	69.93_(1.99)	0.0_(0.01)	76.57_(2.6)	0.08_(0.06)
	Ours (offline, mean)	10.39_(0.62)	0.031_(0.018)	76.77_(1.47)	0.04_(0.03)	69.97_(1.93)	0.0_(0.0)	71.8_(1.59)	0.06_(0.05)
	Ours (FNPM, mean)	10.39_(0.62)	0.031_(0.018)	76.7_(1.43)	0.04_(0.03)	69.97_(1.93)	0.0_(0.0)	72.63_(1.78)	0.08_(0.05)
	Ours (offline, $m$ =10)	6.36_(0.51)	0.017_(0.008)	72.6_(2.37)	0.03_(0.02)	69.97_(1.93)	0.0_(0.0)	71.2_(2.14)	0.02_(0.03)
	Ours (FNPM, $m$ =10)	6.55_(0.44)	0.019_(0.01)	73.0_(2.36)	0.02_(0.02)	69.97_(1.93)	0.0_(0.0)	70.83_(2.25)	0.02_(0.02)
	Ours (offline, $m$ =25)	4.89_(0.24)	0.032_(0.018)	72.77_(2.57)	0.04_(0.04)	69.97_(1.93)	0.0_(0.0)	71.27_(2.72)	0.03_(0.04)
	Ours (FNPM, $m$ =25)	4.91_(0.29)	0.028_(0.018)	72.53_(2.37)	0.03_(0.02)	69.97_(1.93)	0.0_(0.0)	70.9_(2.33)	0.02_(0.02)
10	PCA	38.19_(0.85)	0.137_(0.012)	99.93_(0.13)	0.09_(0.06)	74.43_(0.68)	0.14_(0.14)	95.7_(1.93)	0.11_(0.07)
	Olfat & Aswani, 2019 (0.1)	29.1_(0.98)	0.022_(0.006)	99.97_(0.1)	0.1_(0.06)	71.43_(2.47)	0.03_(0.04)	0.0_(0.0)	0.0_(0.0)
	Olfat & Aswani, 2019 (0.0)	29.0_(0.98)	0.021_(0.005)	99.97_(0.1)	0.1_(0.06)	71.3_(2.44)	0.02_(0.04)	0.0_(0.0)	0.0_(0.0)
	Lee et al., 2022 (1e-3)	33.04_(1.11)	0.022_(0.006)	100.0_(0.0)	0.1_(0.06)	74.3_(1.66)	0.09_(0.02)	95.9_(1.67)	0.09_(0.07)
	Lee et al., 2022 (1e-6)	16.6_(1.08)	0.014_(0.003)	94.43_(1.55)	0.07_(0.06)	69.97_(1.93)	0.0_(0.0)	83.1_(2.93)	0.06_(0.06)
	Kleindessner et al., 2023 (mean)	35.72_(0.83)	0.024_(0.004)	99.97_(0.1)	0.1_(0.06)	74.1_(2.09)	0.09_(0.05)	96.77_(2.16)	0.09_(0.06)
	Kleindessner et al., 2023 (0.85)	27.69_(0.82)	0.02_(0.004)	100.0_(0.0)	0.1_(0.06)	72.97_(3.06)	0.05_(0.05)	97.8_(1.29)	0.09_(0.06)
	Kleindessner et al., 2023 (0.5)	19.87_(0.61)	0.016_(0.004)	99.9_(0.15)	0.1_(0.06)	70.83_(2.87)	0.01_(0.02)	94.67_(2.11)	0.08_(0.05)
	Kleindessner et al., 2023 (kernel)	N/A	N/A	70.1_(1.18)	0.0_(0.01)	69.8_(1.21)	0.0_(0.0)	81.7_(5.49)	0.08_(0.05)
	Ravfogel et al., 2020	15.0_(0.88)	0.01_(0.002)	99.1_(0.47)	0.1_(0.06)	70.7_(2.39)	0.01_(0.02)	93.27_(2.32)	0.07_(0.06)
	Ravfogel et al., 2022	16.41_(1.1)	0.022_(0.011)	99.9_(0.15)	0.1_(0.06)	71.8_(3.08)	0.03_(0.04)	95.9_(2.3)	0.09_(0.04)
	Samadi et al., 2018	N/A	N/A	99.07_(0.57)	0.1_(0.06)	75.87_(0.92)	0.11_(0.1)	98.8_(0.62)	0.09_(0.06)
	Ours (offline, mean)	35.72_(0.83)	0.024_(0.004)	99.97_(0.1)	0.1_(0.06)	74.1_(2.09)	0.09_(0.05)	97.5_(0.85)	0.1_(0.07)
	Ours (FNPM, mean)	35.73_(0.84)	0.024_(0.004)	99.97_(0.1)	0.1_(0.06)	74.07_(2.12)	0.09_(0.05)	95.53_(1.46)	0.1_(0.07)
	Ours (offline, $m$ =10)	26.24_(0.94)	0.018_(0.004)	99.9_(0.21)	0.1_(0.06)	72.83_(3.07)	0.04_(0.04)	96.73_(1.21)	0.1_(0.07)
	Ours (FNPM, $m$ =10)	26.23_(0.81)	0.018_(0.004)	99.9_(0.21)	0.1_(0.06)	71.37_(3.41)	0.02_(0.04)	96.4_(1.33)	0.08_(0.06)
	Ours (offline, $m$ =25)	21.1_(0.48)	0.027_(0.005)	99.87_(0.22)	0.1_(0.06)	71.77_(3.56)	0.02_(0.03)	93.5_(3.54)	0.08_(0.04)
	Ours (FNPM, $m$ =25)	21.06_(0.52)	0.026_(0.005)	99.9_(0.15)	0.1_(0.06)	71.77_(3.41)	0.03_(0.04)	94.97_(2.55)	0.08_(0.05)

Supplemantary for Fair Streaming Principal Component Analysis: Statistical and Algorithmic Viewpoint

1 Synthetic Experiments: Verifying FNPM is PAFO-Learnble

2 Numerical Experiments: Fair PCA on UCI datasets

Supplemantary for
Fair Streaming Principal Component Analysis: Statistical and Algorithmic Viewpoint